Independent Submission S. Das Internet-Draft Independent Intended status: Informational 18 November 2026 Expires: 22 May 2027 AI Safety for AI-Assisted Cyberattacks: Defender-Side Execution Finality at the Protected Consequence Boundary draft-das-ai-safety-agentic-cyberattack-finality-00 Abstract Autonomous AI agents can now carry out parts of a cyber operation that previously needed repeated human effort: reconnaissance, tool selection, execution attempts, observation of results, and adaptation when a route fails. An attacker-controlled agent does not have to malfunction to be dangerous; it can follow malicious instructions efficiently, using stolen credentials and well-formed requests that existing controls accept as valid. This document describes a defender-side architecture, called execution finality, that adds one bounded layer to existing security controls. An incoming request, credential, permitted API call, or agent-composed workflow does not by itself create authority for a sensitive consequence. The defender reconstructs the operation that would actually take effect, holds it in a non-effective state, evaluates authorization independently of the requester, and enforces the result at an enforcement point located at or immediately before the protected consequence. The document specifies placement rules, requirements, state handling, failure behavior, latency strategy, and test methods. It uses publicly reported AI-assisted activity against Taiwanese government agencies in 2026 as an illustrative threat scenario and does not claim that the architecture would have prevented that activity. Status of This Document This note is to be removed before publishing as an RFC. This is an individual, independent-submission Internet-Draft. It is a technical proposal and has not been adopted by any IETF working group. It reports no measured performance, formal proof, or deployment results. Basis of disclosure: this document is based on the international patent application PCT/IB2026/055615, filed 4 June 2026 and publicly disclosed (published as WO 2026/150382 A2) on 16 July 2026. The full filing can be read by searching "Sangam Das" on WIPO PATENTSCOPE, by Das Expires 22 May 2027 [Page 1] Internet-Draft Agentic Cyberattack Execution Finality November 2026 searching PCT/IB2026/055615 (pctib202655615), or at https://www.tdcommons.org/dpubs_series/12542/ . The filing text, not this draft, is authoritative for the patent disclosure. Any mention of third-party products, standards, or organizations is for technical context only. It is not an endorsement, validation, or statement of infringement, and no third-party adoption is implied. Status of This Memo This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79. Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet- Drafts is at https://datatracker.ietf.org/drafts/current/. Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress." This Internet-Draft will expire on 22 May 2027. Copyright Notice Copyright (c) 2026 IETF Trust and the persons identified as the document authors. All rights reserved. This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/ license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Table of Contents 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 4 1.1. Scope . . . . . . . . . . . . . . . . . . . . . . . . . . 5 1.2. Non-Goals . . . . . . . . . . . . . . . . . . . . . . . . 5 2. Terminology . . . . . . . . . . . . . . . . . . . . . . . . . 5 3. Problem Statement and Threat Model . . . . . . . . . . . . . 6 3.1. The Offensive Chain . . . . . . . . . . . . . . . . . . . 6 3.2. Attacker Capabilities . . . . . . . . . . . . . . . . . . 7 3.3. What the Defender Does Not Control . . . . . . . . . . . 7 3.4. Core Invariant . . . . . . . . . . . . . . . . . . . . . 8 4. Architecture Overview . . . . . . . . . . . . . . . . . . . . 8 Das Expires 22 May 2027 [Page 2] Internet-Draft Agentic Cyberattack Execution Finality November 2026 5. Core Principles . . . . . . . . . . . . . . . . . . . . . . . 9 6. The Attacker's AI Need Not Cooperate . . . . . . . . . . . . 10 6.1. The Defender Creates the Candidate Act . . . . . . . . . 10 6.2. Compromise Should Not Propagate Authority . . . . . . . . 10 7. Requirements . . . . . . . . . . . . . . . . . . . . . . . . 11 8. Finality Sink Placement . . . . . . . . . . . . . . . . . . . 12 8.1. Placement Rule . . . . . . . . . . . . . . . . . . . . . 12 8.2. Placement by Consequence . . . . . . . . . . . . . . . . 13 9. Conceptual Data Objects . . . . . . . . . . . . . . . . . . . 14 10. Act States . . . . . . . . . . . . . . . . . . . . . . . . . 14 11. Relationship to Existing Controls . . . . . . . . . . . . . . 15 12. Authority Binding Properties . . . . . . . . . . . . . . . . 17 12.1. Non-Bearer Finality Authority . . . . . . . . . . . . . 17 12.2. Destination Binding . . . . . . . . . . . . . . . . . . 17 12.3. Purpose Binding . . . . . . . . . . . . . . . . . . . . 17 12.4. The Composition Problem . . . . . . . . . . . . . . . . 18 12.5. Aggregate Budgets Against Fragmentation . . . . . . . . 18 12.6. Prepare Is Not Finalize . . . . . . . . . . . . . . . . 18 12.7. The Irreversible Boundary . . . . . . . . . . . . . . . 19 13. Human Oversight, Fail-Closed Behavior, and Audit . . . . . . 19 13.1. Oversight at Machine Scale . . . . . . . . . . . . . . . 19 13.2. Fail-Closed Matters More as Autonomy Increases . . . . . 19 13.3. Why Post-Hoc Audit Is Not Enough . . . . . . . . . . . . 20 14. Control Comparison Matrices . . . . . . . . . . . . . . . . . 20 14.1. Identity and Permission . . . . . . . . . . . . . . . . 20 14.2. Detection and Containment . . . . . . . . . . . . . . . 21 14.3. Protected Data Disclosure Placement . . . . . . . . . . 22 15. Policy Governance, Delegation, and Multi-Organization Use . . 22 15.1. Who Defines the Authorized Envelope . . . . . . . . . . 22 15.2. Agent-to-Agent Delegation . . . . . . . . . . . . . . . 22 15.3. Multiple Organizations and Clouds . . . . . . . . . . . 23 15.4. Conflicting Policies . . . . . . . . . . . . . . . . . . 23 16. Additional Bypass and Residual-Risk Cases . . . . . . . . . . 23 17. Illustrative Scenario: AI-Assisted Activity Against Government Agencies . . . . . . . . . . . . . . . . . . . . . . . . 25 17.1. What Was Publicly Reported . . . . . . . . . . . . . . . 25 17.2. Stolen-Credential Export Example . . . . . . . . . . . . 25 17.3. What the Incident Does and Does Not Show . . . . . . . . 25 18. Accountability and Evidence . . . . . . . . . . . . . . . . . 26 19. Regulatory and Standards Considerations . . . . . . . . . . . 26 19.1. A Technology-Neutral Standards Question . . . . . . . . 27 19.2. From AI Safety to AI Authority . . . . . . . . . . . . . 27 20. Deployment Considerations . . . . . . . . . . . . . . . . . . 27 20.1. Latency and Availability . . . . . . . . . . . . . . . . 27 20.2. Incremental Rollout . . . . . . . . . . . . . . . . . . 27 20.3. Legacy Systems . . . . . . . . . . . . . . . . . . . . . 28 21. Threat Simulation and Mitigation . . . . . . . . . . . . . . 28 21.1. Ground Rules . . . . . . . . . . . . . . . . . . . . . . 28 Das Expires 22 May 2027 [Page 3] Internet-Draft Agentic Cyberattack Execution Finality November 2026 21.2. Simulation Catalogue . . . . . . . . . . . . . . . . . . 28 21.3. Worked Simulation: S1, Stolen Credential Export . . . . 30 21.4. Layered Mitigation Order . . . . . . . . . . . . . . . . 31 21.5. Scoring and Reporting . . . . . . . . . . . . . . . . . 32 22. Evaluation and Testing . . . . . . . . . . . . . . . . . . . 32 23. Limitations . . . . . . . . . . . . . . . . . . . . . . . . . 33 24. Security Considerations . . . . . . . . . . . . . . . . . . . 33 25. Privacy Considerations . . . . . . . . . . . . . . . . . . . 34 26. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 34 27. References . . . . . . . . . . . . . . . . . . . . . . . . . 34 27.1. Normative References . . . . . . . . . . . . . . . . . . 34 27.2. Informative References . . . . . . . . . . . . . . . . . 34 Appendix A. Frequently Asked Questions . . . . . . . . . . . . . 35 A.1. Basics . . . . . . . . . . . . . . . . . . . . . . . . . 35 A.2. Engineering . . . . . . . . . . . . . . . . . . . . . . . 38 A.3. Cross-organization and delegation . . . . . . . . . . . . 41 A.4. Evidence, testing, and claims . . . . . . . . . . . . . . 41 Basis of Disclosure and Patent Notice . . . . . . . . . . . . . . 43 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . 43 Author's Address . . . . . . . . . . . . . . . . . . . . . . . . 43 1. Introduction Much AI safety work addresses accidental failure: hallucination, misalignment, or unintended tool use. A separate problem is deliberate use. An attacker supplies the objective, and AI agents perform substantial parts of the work: finding targets, choosing tools, trying routes, reading results, and trying another route when one is blocked. On 13 August 2026, Taiwan's Ministry of Digital Affairs publicly confirmed that government agencies had been targeted during July 2026 by an overseas campaign using both conventional techniques and AI agents, and stated that its defenses detected the abnormal activity and that the affected agencies handled the incident [TAIWAN-MODA]. The background rate is also high: Taiwan's National Security Bureau reported an average of roughly 2.63 million intrusion attempts per day against its infrastructure in 2025 [TAIWAN-NSB]. Public reporting does not establish the intrusion path, the protected resources involved, or what would have happened under any alternative defense. This document therefore treats the incident only as an illustrative scenario (Section 17). Das Expires 22 May 2027 [Page 4] Internet-Draft Agentic Cyberattack Execution Finality November 2026 The lesson drawn here is narrow. An AI agent need not "go rogue." It can follow hostile instructions competently. When it does, the target system often sees a valid credential, a well-formed request, and a permitted API call. Authentication and coarse authorization answer whether a caller may use a function or resource. They do not always answer whether this specific operation, with this final payload and destination, may take effect now. This document describes a complementary layer that asks that second question at the point of consequence. It does not replace identity and access management, zero trust architectures, endpoint detection, network security, or transactional controls. It identifies where those controls may leave a gap and specifies what an additional enforcement layer needs in order to close it. 1.1. Scope This document covers defender-controlled systems that can commit, release, transmit, or actuate a protected effect, for example a data export, a configuration change, a secret release, a financial posting, or an actuator command. It does not attempt to control, detect, or attribute the attacker's AI. It does not address model training, alignment, or content moderation. 1.2. Non-Goals * Preventing an attacker from using AI for reconnaissance or planning. * Replacing detection, response, patching, or segmentation. * Defending against full compromise of the enforcement layer itself (see Section 23). * Making claims about attribution of any incident. 2. Terminology The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here. The names below describe architectural roles in this proposal. They are not established protocol or standard terms. Das Expires 22 May 2027 [Page 5] Internet-Draft Agentic Cyberattack Execution Finality November 2026 Protected Consequence An effect whose unauthorized occurrence is unacceptable: an external data release, a privileged change, a key or secret release, a financial posting, or a physical actuation. Candidate Act The defender-side representation of the operation that would actually take effect if allowed, reconstructed from the incoming request and system state rather than taken from the requester's description. Non-Effective State A state in which the Candidate Act is staged but has produced no protected effect. Protected Validator A defender-controlled component, isolated from the requesting workload, that evaluates authorization predicates for a Candidate Act. Finality Sink The defender-controlled enforcement point, at or immediately before the protected consequence, that permits the effect only if bounded, current, act-specific authority is present. Bounded Authority Authorization limited to a specific act, destination, quantity, time window, and policy epoch, and consumable at most the permitted number of times. Receipt Authenticated evidence, produced by the Finality Sink, of the outcome (committed, denied, held, or indeterminate) for one Candidate Act. 3. Problem Statement and Threat Model 3.1. The Offensive Chain A traditional operation often needs a human to inspect each result and decide the next step. An AI-assisted operation can compress that loop: Das Expires 22 May 2027 [Page 6] Internet-Draft Agentic Cyberattack Execution Finality November 2026 Human objective -> one or more AI agents -> reconnaissance -> target and vulnerability analysis -> tool selection -> attempt -> observe response -> re-plan / try another path -> attempted external consequence The defender's problem is not that the agent has an unusual intent. It is that adaptation is cheap, retries are fast, and each attempt can look like ordinary authenticated traffic. 3.2. Attacker Capabilities The attacker MAY: * operate AI agents the defender cannot inspect, constrain, or influence; * hold stolen or misused credentials, tokens, or service identities; * send syntactically valid, authenticated requests through ordinary interfaces; * inject instructions into content, tool output, or delegated sub- agents the defender's own AI processes; * try alternative routes (direct APIs, queues, batch jobs, admin paths) when one is blocked; * claim in natural language that an operation is authorized. The model assumes the attacker does not simultaneously compromise the Finality Sink, the Protected Validator, the authority store, and all audit and reconciliation channels. See Section 23. 3.3. What the Defender Does Not Control The defender cannot require the attacker's AI to use any guardrail, logging, identity scheme, or policy. Controls placed inside the attacker's agent are therefore out of scope. The enforcement opportunity begins when a request reaches a defender-controlled system. Das Expires 22 May 2027 [Page 7] Internet-Draft Agentic Cyberattack Execution Finality November 2026 3.4. Core Invariant For a Protected Consequence, the existence of a machine-generated request, a valid credential, a technically permitted API call, or a statement of authorization MUST NOT by itself cause the effect. The effect occurs only when bounded authority bound to the exact Candidate Act is verified at the Finality Sink. Corollary: an agent that discovers a new route, tool, or vulnerability MUST NOT thereby gain authority over any additional Protected Consequence. Adaptation changes how an attempt arrives. It does not change what is authorized. 4. Architecture Overview ATTACKER DOMAIN | DEFENDER TRUST BOUNDARY | Hostile AI / | stolen credential | | | | ordinary | | request | +--------------->| Protected service / API | | | v | Reconstruct Candidate Act | | | v | Non-Effective State (staged) | | | v | Protected Validator | (act, destination, purpose, | quantity, time, policy epoch, | revocation, cumulative exposure) | | | +----+---------------+ | | | | deny / hold / issue Bounded | reconcile Authority | | | | v v | NO EFFECT Finality Sink | (fail closed) verifies act-bound, | current authority | | | v | PROTECTED EFFECT + Receipt Das Expires 22 May 2027 [Page 8] Internet-Draft Agentic Cyberattack Execution Finality November 2026 Figure 1: Defender-side execution-finality pipeline The Candidate Act is created on the defender side. No cooperation from the attacker or the attacker's AI is assumed. The Protected Validator and Finality Sink SHOULD be separated from the requesting workload so that compromise of the workload does not directly yield authority. 5. Core Principles The architecture reduces to four principles. P1. Computation is not authority. A machine-generated request, decision, or command does not itself establish authority for the protected system to make it effective. The defender cannot stop a hostile AI from computing a malicious action. It can ensure the computation alone has no effect. P2. Adaptation does not expand authority. An agent that finds another route, tool, vulnerability, or strategy does not thereby enlarge its authority over protected consequences. Path A denied, then Path B, then Path C: each path that reaches a protected consequence faces the same authorization conditions. P3. Control consequence, not the attacker's intelligence. The defender may never control what the hostile AI thinks or what strategy it generates next. It can seek stronger control over what its own protected infrastructure is permitted to make effective: its data, money, communications, infrastructure, and physical systems at the point they cross from internal state to external effect. P4. Evidence precedes effect. For high-consequence operations, authorization evidence SHOULD exist before, or atomically with, permission to effectuate. It SHOULD NOT exist only as an audit trail created afterwards. A summary of the invariant: machine decision is not machine authority. The AI may reason, adapt, discover, and request. Authority to turn computation into consequence stays independently bounded, verifiable, and enforceable at the protected consequence boundary. Das Expires 22 May 2027 [Page 9] Internet-Draft Agentic Cyberattack Execution Finality November 2026 6. The Attacker's AI Need Not Cooperate A common misunderstanding is that this architecture needs the attacking AI to run inside, or take part in, the protected system. It does not. Consider a hostile AI running entirely on infrastructure outside the target organization. It learns that an employee account can reach a sensitive government database, obtains valid credentials, and sends an apparently ordinary request such as "retrieve records matching these identifiers." The request arrives over an encrypted, authenticated channel through a legitimate API, and the application holds database permission. Traditional authentication did its job. None of those facts establishes that the resulting consequence is authorized. 6.1. The Defender Creates the Candidate Act The attacker does not need to send anything called a Candidate Act and may know nothing about the architecture. The protected system interprets the requested consequential operation as a Candidate Act before allowing the effect: External request -> application receives request -> application proposes "release protected records" -> protected architecture treats the proposal as a Candidate Act -> Candidate Act remains NON-EFFECTIVE -> protected validation -> finality decision The security property resides with the defender. The attacker cannot avoid the gate by declining to participate, provided the protected consequence cannot be reached through another unmediated path. This is analogous to a building's access-control system: a visitor does not need to implement it, because the building does. 6.2. Compromise Should Not Propagate Authority Traditional compromise often creates a chain: compromise an account, inherit its permissions, use application privileges, reach data, reach the network, create a consequence. Agentic AI makes this worse because the attacker can automate the search for ways to exploit inherited privileges. The architecture attempts to interrupt that propagation: * Compromise does not imply finality authority. Das Expires 22 May 2027 [Page 10] Internet-Draft Agentic Cyberattack Execution Finality November 2026 * Authentication does not imply finality authority. * Application permission does not imply finality authority. * Tool access does not imply finality authority. * Network access does not imply finality authority. Invariant: compromise of an upstream authority SHOULD NOT automatically propagate into unrestricted authority over downstream consequences. 7. Requirements R1. Actual-operation reconstruction. The system MUST validate the operation that will really execute, not the API name or the requester's description. It MUST account for indirect effects: triggers, stored procedures, fan-out, batch jobs, delegated services, redirects, and downstream recipients. R2. Authorization predicates. The Protected Validator MUST evaluate, as applicable: workload and requesting authority, act identity, operation, data classification, destination, purpose, quantity, time, environment health, policy version, revocation epoch, and cumulative exposure. Claims made in the request MUST be treated as evidence to evaluate, not as policy. R3. Act binding and non-bearer authority. Bounded Authority MUST be bound to the specific act and destination, a protected workload identity, freshness (nonce or counter), policy epoch, and sink-local one-time or bounded consumption. A copied token or receipt MUST NOT authorize a different effect. R4. Independent enforcement. The Protected Validator and Finality Sink SHOULD be separated from the compromised-workload trust domain wherever feasible. Policy updates, key material, and decision state MUST be protected. Relevant components SHOULD be attested where appropriate. Das Expires 22 May 2027 [Page 11] Internet-Draft Agentic Cyberattack Execution Finality November 2026 R5. Cross-path closure. Deployments MUST enumerate every route to the Protected Consequence (direct API, internal service, asynchronous queue, database administration, backup or export, logs, browser, email, cloud storage, privileged batch, secondary network interface, recovery interface). Any route that is not mediated MUST be recorded as residual risk. The deployment MUST NOT claim non- bypassability for such routes. R6. Race and atomicity. The decision MUST be bound to the operation version and current destination and security state. Changed conditions MUST trigger revalidation. For writes, the check SHOULD be coordinated with transaction commit. For transmissions, the release itself SHOULD be controlled. R7. Partial effect handling. Where an effect can partially occur, the system MUST define idempotent recovery, reconciliation, and compensating actions. It MUST NOT assume a clean rollback. R8. Fail closed, with nuance. If required authority is absent, stale, or unverifiable, the Finality Sink MUST prevent the Protected Consequence. For ambiguous partial completion, the system SHOULD HOLD and RECONCILE rather than retry blindly. For safety-critical physical systems, predefined safe-state or emergency paths MUST be engineered and authorized separately, because indiscriminate blocking can itself cause harm. R9. Evidence before effect. For high-consequence operations, the authorization evidence SHOULD exist before, or atomically with, the effect rather than being reconstructed from logs afterwards. 8. Finality Sink Placement 8.1. Placement Rule The Finality Sink SHOULD be located in the lowest practical trusted component that exclusively controls the Protected Consequence, and all equivalent paths MUST converge on it. A check placed higher in the stack is supplementary, not final, if an alternate API, batch job, privileged account, internal service, or direct driver can reach the same effect without passing through it. Das Expires 22 May 2027 [Page 12] Internet-Draft Agentic Cyberattack Execution Finality November 2026 An API gateway or agent dispatcher that approves a request but leaves downstream workers, direct database connections, alternate cloud APIs, scheduled tasks, or maintenance routes able to produce the same result is not a complete boundary. In distributed systems there MAY be several cooperating sinks, each governing its own consequential transition. 8.2. Placement by Consequence +=================+================================================+ | Protected | Sink location | | effect | | +=================+================================================+ | Sensitive | Protected query-result release or export | | database export | service, before records leave the trusted data | | | domain; paired with egress controls | +-----------------+------------------------------------------------+ | Database | Database transaction commit, privileged write | | mutation or | service, or enforced stored-program effect | | deletion | boundary | +-----------------+------------------------------------------------+ | File or object | Object-store retrieval or release gateway, or | | disclosure | protected key release, before plaintext is | | | externally usable | +-----------------+------------------------------------------------+ | Network | Exclusive egress gateway or protected | | exfiltration | transmission release boundary, coordinated | | | with data-origin controls | +-----------------+------------------------------------------------+ | Cloud | Privileged cloud-control API execution proxy | | administration | or enforcement service, before the resource or | | | configuration change | +-----------------+------------------------------------------------+ | Credentials and | HSM- or KMS-backed secret release, issuance, | | secrets | signing, or privileged credential activation | +-----------------+------------------------------------------------+ | MCP and agent | Destination tool-effect gateway or protected | | tool workflows | service; an MCP proxy alone is insufficient if | | | direct access remains | +-----------------+------------------------------------------------+ | Financial | Capture, payment authorization, ledger | | transfer | posting, or settlement execution boundary | | | appropriate to the effect | +-----------------+------------------------------------------------+ | Industrial and | Trusted actuator command gate, safety | | robotics | controller, or protected physical output | | | interface | +-----------------+------------------------------------------------+ Das Expires 22 May 2027 [Page 13] Internet-Draft Agentic Cyberattack Execution Finality November 2026 | Telecom and 6G | Configuration commit, privileged control-plane | | networks | transition, or transmission authorization; PHY | | | and HARQ timing loops are avoided | +-----------------+------------------------------------------------+ Table 1: Recommended sink location by protected effect 9. Conceptual Data Objects This section describes the information each object needs. It does not define a wire format. A future revision, or a profile for a specific protocol, could define serialization, canonicalization, and signature suites. Candidate Act descriptor Act identifier; resolved operation and target; final payload digest; data classification and quantity; recipient or destination identity; requesting workload identity; object or resource version; timestamps. Bounded Authority Reference to the Candidate Act descriptor digest; issuer identity; policy epoch; revocation epoch; validity window; maximum consumption count; sink identity; nonce or counter; issuer signature or equivalent authenticator. Receipt Act identifier; Bounded Authority reference; outcome (COMMITTED, DENIED, HELD, INDETERMINATE, RECONCILED); sink identity; sink- local counter; timestamp; authenticator. A Receipt MAY be required before a later phase of a multi-step operation proceeds. Implementers can build much of this from existing primitives, such as signed tokens, hash binding, and fine-grained authorization details [RFC9396], token exchange [RFC8693], sender-constrained tokens [RFC9449], and remote attestation [RFC9334]. This document does not claim novelty for any of them individually. The point of the architecture is the integrated invariants in Section 7. 10. Act States An implementation SHOULD track each Candidate Act through the following states. NON-EFFECTIVE Staged; no protected effect. Das Expires 22 May 2027 [Page 14] Internet-Draft Agentic Cyberattack Execution Finality November 2026 HELD Awaiting validation, additional evidence, or revalidation after changed conditions. AUTHORIZED Bounded Authority issued but not yet consumed. COMMITTED The sink consumed the authority and the effect occurred; Receipt issued. DENIED Validation failed or authority was absent, stale, or mismatched; no effect. INDETERMINATE Outcome unknown, for example after a crash or timeout between sink consumption and confirmation. RECONCILED Outcome established by querying authoritative destination state; retries are idempotent against that state. Consumption of Bounded Authority and the effect SHOULD be atomic where the substrate allows. Where it cannot be, the sink MUST persist consumption durably before the effect and MUST resolve INDETERMINATE acts by reconciliation, not by re-issuing authority. 11. Relationship to Existing Controls Existing controls remain necessary. The comparison below is a shorthand for where an additional layer may add value, not a claim that existing controls cannot do these things. A well-engineered policy enforcement point under attribute-based access control or zero trust [NIST-SP-800-207] can already evaluate action-specific attributes and may implement much of what is described here. Das Expires 22 May 2027 [Page 15] Internet-Draft Agentic Cyberattack Execution Finality November 2026 +================+====================+=============================+ | Control | What it provides | Remaining question | +================+====================+=============================+ | Authentication | Establishes who or | Is this specific | | | what is calling | consequence authorized, | | | | even for a valid caller? | +----------------+--------------------+-----------------------------+ | IAM, RBAC, | Resource- and | Is the enforcement point | | ABAC | attribute-level | at the actual effect, | | | permission | and does it bind the | | | | final payload? | +----------------+--------------------+-----------------------------+ | OAuth tokens | Scoped and sender- | Does the sink verify the | | and | constrained | final effective payload, | | authorization | delegation | target, and version, and | | details | | the commit outcome? | +----------------+--------------------+-----------------------------+ | Sandboxing | Contains a process | What if the legitimate | | | | environment already | | | | exposes a harmful | | | | capability? | +----------------+--------------------+-----------------------------+ | Firewall, IDS/ | Detect and | If an action is | | IPS, EDR, zero | constrain | technically possible and | | trust | | reaches the protected | | | | system, does it hold | | | | bounded authority? | +----------------+--------------------+-----------------------------+ | Model | Reduce unsafe | Cannot be the final | | guardrails | model output | boundary against an | | | | attacker who controls | | | | the model or prompt | +----------------+--------------------+-----------------------------+ | Logging and | Evidence after the | Can authorization | | audit | effect | evidence exist before | | | | the effect? | +----------------+--------------------+-----------------------------+ Table 2: Existing controls and the remaining question For tool-using agents, an agent-to-tool dispatch policy, for example at an MCP client or proxy, is useful but advisory. The enforcement point that matters is the tool server's effect executor or the downstream target API or credential broker. Tool output MUST NOT be treated as proof that an effect was committed. For agent-to-agent handoff, parent identity and task metadata do not by themselves grant new power; a child action needs authority issued by a trusted delegator that attenuates the parent's restrictions. Das Expires 22 May 2027 [Page 16] Internet-Draft Agentic Cyberattack Execution Finality November 2026 12. Authority Binding Properties 12.1. Non-Bearer Finality Authority If execution authority were an ordinary transferable token, a compromised application could steal and replay it, recreating the original problem one layer down: steal a credential to gain access becomes steal a finality token to gain effectuation authority. Authority SHOULD therefore not behave as a bearer credential whose possession alone suffices. It can instead be tied to protected state for the intended consequence, which may include protected workload identity, Candidate Act attributes, purpose, resource, destination, security epoch, freshness state, protected execution environment, and sink-local consumption state. Copying an authorization artifact SHOULD NOT reproduce the authority needed to effectuate the consequence. Stealing evidence of authority is not the same as stealing the authority. 12.2. Destination Binding Destination control is especially important against AI-assisted attacks. A legitimate government application may normally export reports to an authorized repository. A compromised process may instead try to send sensitive records to an external endpoint. At the coarse level of "the application has network access" both look alike, but the consequences differ completely. Destination SHOULD be part of the authorization decision: authorized resource, purpose, and destination pass; an authorized resource with an unauthorized destination fails. Broad network permission is not consequence- specific authority. 12.3. Purpose Binding An employee or AI assistant may legitimately access personnel data for payroll administration. That does not imply authority to use the same data for external profiling, unrelated analytics, disclosure, model training, or export. Conventional access control sees an authorized user and an authorized database. A finality check can additionally evaluate authorized purpose and authorized consequence. This matters in agentic settings because AI can combine individually legitimate capabilities in ways nobody anticipated when the permissions were granted. Das Expires 22 May 2027 [Page 17] Internet-Draft Agentic Cyberattack Execution Finality November 2026 12.4. The Composition Problem Suppose an agent holds three individually authorized capabilities: read internal documents, summarize information, send email. The agent reads a confidential document, extracts sensitive information, summarizes it, and sends the summary to an unauthorized external recipient. Every tool operated correctly. The harm arises from the composition, which exposes a limit of tool-by-tool permission models. The question cannot always stop at whether the AI was permitted to call the send tool. It may need to be whether the consequence produced by this sequence was authorized. The final external disclosure SHOULD be treated as the protected Candidate Act. The chain of tool calls ends in a candidate disclosure that stays NON-EFFECTIVE until validation confirms destination, purpose, and scope. For multi-step workflows, consequential steps SHOULD be authorized by their own effects, and later authority SHOULD be bound to verified earlier outcomes where necessary. A workflow- level permission MUST NOT become blanket permission for every later tool and destination. 12.5. Aggregate Budgets Against Fragmentation An attacker can split one large export into thousands of small acts that each pass a per-act limit. Implementations SHOULD enforce aggregate budgets across time windows, destinations, identities, and related operations, using protected accounting that cannot be reset by changing tools or sessions. They MUST account for parallel requests and races so that simultaneous small transfers cannot together exceed the authorized exposure budget. 12.6. Prepare Is Not Finalize Tool invocation authority is not the same as consequence finality authority. An AI MAY be permitted to prepare an operation without being permitted to finalize it: * prepare a payment, without settlement authority; * prepare an external message, without transmission authority; * prepare a database modification, without commit authority; * prepare an industrial command, without actuation authority; * prepare a sensitive-data export, without release authority. Das Expires 22 May 2027 [Page 18] Internet-Draft Agentic Cyberattack Execution Finality November 2026 The intelligence layer decides what to do. The protected finality layer decides whether this particular consequence may occur. Finality controls are targeted at the transition where a proposed action becomes consequential, not at every internal computation. 12.7. The Irreversible Boundary Not every computation needs finality enforcement. The architecture matters most near the point where an action becomes externally usable, externally visible, financially settled, physically actuated, transmitted, committed, disclosed, or otherwise hard to reverse. Drafting an email is not sending it. Calculating a payment is not settling it. Generating an industrial command is not moving the actuator. Preparing a database update is not committing it. Reading sensitive data internally is not necessarily releasing it externally. Staging itself MUST NOT leak the effect indirectly, because temporary files, logs, telemetry, caches, and previews can become release channels. 13. Human Oversight, Fail-Closed Behavior, and Audit 13.1. Oversight at Machine Scale Human approval of every action does not scale when autonomous systems produce hundreds, thousands, or millions of machine-speed decisions. Granting broad standing authority to avoid constant approvals creates excessive risk. A third option is machine-verifiable bounded authorization: humans or organizational policy define the authorized envelope; machines generate Candidate Acts; protected mechanisms check each act against the envelope at effectuation time; humans intervene where policy requires escalation. This does not remove human oversight. It moves human judgment to defining authority and handling escalations. 13.2. Fail-Closed Matters More as Autonomy Increases A permissive system that executes when authorization is missing, stale, ambiguous, unverifiable, revoked, inconsistent, or unavailable, and investigates afterwards, is dangerous for high- consequence autonomous systems. A human operator may make one questionable decision. An autonomous system may repeat a mistaken assumption across thousands of operations before anyone notices. Fail-closed enforcement limits the ability of uncertainty to propagate into real-world consequences. The safety-critical qualification in Section 7 R8 still applies. Das Expires 22 May 2027 [Page 19] Internet-Draft Agentic Cyberattack Execution Finality November 2026 13.3. Why Post-Hoc Audit Is Not Enough A perfect log of an unauthorized action does not prevent it. After money has moved, data has escaped, an actuator has moved, or a command has been transmitted, excellent audit evidence is valuable but insufficient. Traditional audit runs effect then evidence. Execution finality runs authorization evidence then effect. This does not remove audit; it changes the time relationship between evidence and consequence. For high-consequence machine actions, the strongest evidence is evidence that had to exist before the system permitted the consequence. 14. Control Comparison Matrices The finality layer MUST be evaluated against strong configurations of existing controls, not against a straw-man perimeter firewall. Modern security already includes transaction-level authorization, data-loss prevention, egress controls, database activity monitoring, and hardware isolation. The defensible claim is narrower than "conventional systems fail": existing controls can be correctly implemented and still leave a gap between permission to invoke an interface and authorization for its specific downstream consequence. 14.1. Identity and Permission Authentication verifies an identity claim or credential possession. MFA, device posture, and phishing-resistant authenticators materially reduce many credential attacks, but do not alone prove that an authenticated request is within an authorized purpose, recipient, volume, or workflow. IAM, RBAC, and ABAC can express richer constraints, but deployments may be coarse or inconsistent. OAuth and scoped API access may already support fine-grained resource indicators, audiences, and consent, and a finality design MUST NOT claim these are inherently incapable of consequence control. Its distinctive target is enforcement at the actual side-effecting sink, with a reconstructed effective operation, freshness, non-transferable authorization, and denial if the sink cannot verify it. Das Expires 22 May 2027 [Page 20] Internet-Draft Agentic Cyberattack Execution Finality November 2026 +===================+==================+============================+ | Control | Baseline issue | Proposed test or property | +===================+==================+============================+ | Authentication | Valid credential | Act-specific authorization | +-------------------+------------------+----------------------------+ | IAM / RBAC / | Scope may be | Destination and purpose | | ABAC | coarse | binding | +-------------------+------------------+----------------------------+ | OAuth | API invocation | Effect and revocation | | | scope | check | +-------------------+------------------+----------------------------+ Table 3: Identity and permission controls Test: valid stolen identity with a prohibited destination, comparing strong ABAC policies with and without the sink, including direct- service and queued paths. 14.2. Detection and Containment Firewalls, segmentation, IDS/IPS, EDR, DLP, database controls, and zero trust are complementary. They can block many unauthorized operations before any finality gate is reached. Detection-oriented systems can also miss novel or blended behavior, and prevention- oriented controls may already provide equivalent effect-specific denial. The finality layer SHOULD reuse their telemetry and not duplicate it. Sandboxing is valuable when it restricts reachable resources and system calls, but a compromised process can still abuse a legitimately exposed API. The relevant question is whether every protected effect path is mediated, including service identities, background jobs, replication, bulk exports, and maintenance interfaces. An asserted universal gate without an exhaustive path inventory is not a complete claim. +===========+====================+===========================+ | Control | Baseline issue | Proposed test or property | +===========+====================+===========================+ | EDR / IDS | Detects suspicious | Deny the protected effect | | | behavior | at the sink | +-----------+--------------------+---------------------------+ | DLP / | May already | Test incremental benefit | | egress | prevent release | | +-----------+--------------------+---------------------------+ | Sandbox | Constrains | Mediate legitimate effect | | | reachable tools | paths | +-----------+--------------------+---------------------------+ Table 4: Detection and containment controls Das Expires 22 May 2027 [Page 21] Internet-Draft Agentic Cyberattack Execution Finality November 2026 14.3. Protected Data Disclosure Placement External AI / compromised credential | v Defender application: candidate export held | v Independent validation of scope, recipient, freshness | v FINALITY SINK: protected release / egress / key gate | v Authorized destination or DENY Figure 2: Placement for protected data disclosure 15. Policy Governance, Delegation, and Multi-Organization Use 15.1. Who Defines the Authorized Envelope The statement that an action must fall inside an authorized envelope does not say who defines the envelope, who may expand it, who may revoke it, or what protects those operations. Policy creation is part of the trust model. If policy is too static, legitimate new workflows are blocked and pressure builds to add broad exceptions or disable the control. If policy can be widened silently, authority can expand until the finality layer is meaningless. Mature deployments therefore need controlled policy evolution: changes to the scope of authority are themselves protected acts, authenticated, change-controlled, and auditable. Different environments MAY choose static, periodically reviewed, or context-narrowed envelopes. A policy update MUST NOT retroactively authorize a pending operation unless re-evaluation under the new policy is explicit and the applicable epoch is recorded. During an incident, relevant envelopes SHOULD be frozen or revoked and queued operations prevented from draining under stale approvals. 15.2. Agent-to-Agent Delegation Delegating work is not delegating finality authority. It MUST NOT be the case that Agent B inherits every authority Agent A possessed. Delegated authority SHOULD be attenuated, bounded, and verifiable at the consequence boundary that Agent B's action reaches. Das Expires 22 May 2027 [Page 22] Internet-Draft Agentic Cyberattack Execution Finality November 2026 15.3. Multiple Organizations and Clouds The simplest model assumes one defender controls the agent, the policy, the protected resource, and the consequence boundary. Real workflows often cross organizations. Open questions include whose policy governs the action, which organization issues authority, which verifies finality, whether one organization can trust another's authorization evidence, how authority is revoked across domains, and what happens when jurisdictional rules conflict. Each organization controls its own consequence boundaries. An authorization valid in one domain MUST NOT be presumed valid in another, and the receiving sink MUST independently enforce its own policy. Bounded authority that crosses an organizational boundary MUST NOT turn into a transferable bearer permission that can be replayed or broadened downstream. Cross-domain finality needs explicit protocol and trust- model design (trust anchors, delegated scope, policy composition, authenticated evidence, revocation semantics) and MUST NOT be assumed to work automatically. 15.4. Conflicting Policies When two organizations' policies conflict, a deterministic composition rule is needed. The system MUST NOT simply choose whichever rule allows execution. Acceptable rules include explicit contractual priority, jurisdiction-specific policy, or a fail-closed conflict rule. This is an interoperability matter, not only an implementation detail. 16. Additional Bypass and Residual-Risk Cases Malicious or compromised administrator A privileged account may alter policy, change routing, disable monitoring, or read raw storage. The protection holds only to the extent administrative operations are themselves mediated and protected by stronger controls. Policy administration SHOULD be separated from enforcement, change control authenticated, signing keys protected, exceptional paths audited, and break-glass procedures tested. A software-only gate MUST NOT be described as immune to host-root compromise. Das Expires 22 May 2027 [Page 23] Internet-Draft Agentic Cyberattack Execution Finality November 2026 Read as the irreversible disclosure The consequence may arise when plaintext is delivered to an untrusted process, not only when it later leaves the network. In that case enforcement belongs before the read result is exposed: an exclusive database view, a trusted query broker, or a decryption or key-release boundary. A design that allows unrestricted plaintext reads and checks only later network export cannot guarantee confidentiality against local capture or covert channels. Covert channels and screenshots A release gate cannot guarantee containment if a compromised process can view plaintext and encode it into screenshots, logs, timing signals, or permitted outbound traffic. Sensitive data SHOULD be protected before plaintext exposure, untrusted execution constrained, and DLP and monitoring used. Residual risk MUST be stated, not hidden. Firewall or outbound proxy as sink A firewall can be one enforcement component, but network metadata may not reveal which protected records, purpose, or quantity a TLS connection carries. Decisions about sensitive data may need to occur at the application or data-release layer, with network egress as a complementary control. If an application can send data over an alternate tunnel or local interface, the egress control is not exclusive. Prompt injection Prompt injection can still influence what an agent requests. An independent sink does not accept model-generated text as authorization merely because the agent claims a task is approved. If untrusted prompt content can rewrite policy or trusted attributes, the intended independence has failed. Stolen OAuth or API tokens A stolen token may still permit reconnaissance, reads, or operations outside the finality boundary. The proposal narrows specific protected consequences; it does not make stolen tokens harmless. Least privilege, token binding, monitoring, and rapid revocation remain necessary. Validation-service exhaustion Admission control, rate limits, bounded queues, local verification where feasible, workload isolation, and separate resource budgets for high-priority services SHOULD be used. A finality service that becomes a denial-of-service choke point reduces availability, so resilience is a deployment requirement, not an assumed property. Das Expires 22 May 2027 [Page 24] Internet-Draft Agentic Cyberattack Execution Finality November 2026 Centralization The Finality Sink is a high-value target and MUST NOT be assumed to be one software process on one server acting as universal authority. Designs can use hardware-backed isolation, measured or attested software state, multiple independent trust domains, or separate sinks for separate consequence classes. The architectural principle is a protected boundary, not one centralized component. 17. Illustrative Scenario: AI-Assisted Activity Against Government Agencies 17.1. What Was Publicly Reported Taiwan's Ministry of Digital Affairs reported on 13 August 2026 that government agencies were targeted in July 2026 by an overseas campaign using conventional techniques and AI agents, and that its cybersecurity infrastructure detected the activity and the agencies handled it [TAIWAN-MODA]. This document does not attribute the activity and takes no position on its origin. 17.2. Stolen-Credential Export Example The following is a hypothetical walk-through, not a description of the reported incident. A hostile agent sends a valid, authenticated request to export government records. The application can read the data, but the export remains staged. The protected release service reconstructs the true record set, recipient, and quantity. The Protected Validator evaluates destination, purpose, policy, and revocation. The Finality Sink checks act-bound authority and permits only the exact bounded release. A mismatched external destination is denied even though the stolen credential authenticates. This outcome holds only if the attacker cannot use an unmediated alternative route and cannot compromise the enforcement layer (Section 23). 17.3. What the Incident Does and Does Not Show The incident illustrates that AI can lower the cost of adaptation and parallelism in an operation. It does not show that this architecture would have prevented it, and it is not evidence of deployment or effectiveness in any Taiwanese system. It supports a general point: defenses that depend only on detecting or limiting the attacker's methods face a faster adversary, so it is worth also controlling which consequences can occur at all. Das Expires 22 May 2027 [Page 25] Internet-Draft Agentic Cyberattack Execution Finality November 2026 18. Accountability and Evidence Agentic AI complicates accountability. After an autonomous system executes an important command, investigators ask who gave the objective, which AI generated the action, which account and credential were used, and which application invoked the tool. A further question becomes important: which protected authority actually permitted this specific act to cross into consequence? Traditional logs establish what happened. Finality evidence could additionally establish what authorization state existed before the protected consequence was permitted. That distinction may help incident investigation, organizational accountability, regulatory supervision, contractual responsibility, human oversight, and technical assurance. Such evidence does not itself determine legal liability, but it can give a technically stronger authorization chain. Investigators could then ask both "which AI produced the command?" and "which protected authority permitted this command to cross the finality boundary?" 19. Regulatory and Standards Considerations This section is a technical observation, not a legal analysis. Nothing here claims that any current law mandates this architecture. The European Union's AI Act establishes a risk-based framework that includes requirements concerning areas such as risk management, human oversight, robustness, and cybersecurity for relevant systems. The Cyber Resilience Act sets cybersecurity requirements for products with digital elements across their lifecycle. As implementation moves toward harmonised standards and technical specifications, an additional technical question arises for agentic systems that interact with critical infrastructure, financial systems, telecommunications, public administration, industrial systems, healthcare infrastructure, or transportation: how can authorization be demonstrated at the moment an autonomous machine-generated act becomes an externally effective consequence? A system can have excellent documentation, strong model evaluation, high-quality training data, and detailed audit logs, and still lack a technical mechanism that independently determines whether a specific machine-generated consequence was authorized immediately before effectuation. Execution finality addresses that narrower gap. It is proposed as a technical direction for future standards, critical- infrastructure deployments, and high-consequence agentic systems. Das Expires 22 May 2027 [Page 26] Internet-Draft Agentic Cyberattack Execution Finality November 2026 19.1. A Technology-Neutral Standards Question Rather than mandating one architecture, standards bodies could consider a technology-neutral property: can the system demonstrate that a defined class of high-consequence machine-generated acts cannot become externally effective unless required authorization conditions have been independently satisfied at the relevant consequence boundary? That formulation does not mandate any one token format, hardware architecture, cryptographic scheme, AI model, operating system, or implementation. It defines a security property: no protected consequence without valid authority. The objective is to standardize, where justified, that invariant rather than any one inventor's terminology. 19.2. From AI Safety to AI Authority AI governance has concentrated on what the AI knows, outputs, and can be manipulated into, whether it is aligned, and whether its decisions can be explained. Agentic AI adds a question of authority: what is the AI actually allowed to make happen? An AI that recommends a payment is one risk, and an AI that settles it is another. The same holds for recommending versus actuating a process change, and for identifying versus transmitting sensitive information. As AI moves from information systems into action systems, governance architecture needs to move with it. 20. Deployment Considerations 20.1. Latency and Availability Implementations SHOULD NOT insert human review or heavyweight remote cryptography into every routine operation. Practical approaches include precomputed bounded policy corridors, local deterministic sink checks, cached authenticated state where safe, and stronger checks reserved for consequential transitions. Deployments SHOULD measure p50, p95, and p99 latency, throughput, failover behavior, queue saturation, and denial-of-service behavior. No measured performance is asserted in this document. 20.2. Incremental Rollout Deployments can start with the highest-consequence effects (bulk export, privileged administration, secret release), run the Protected Validator in observe-only mode to compare its decisions with actual outcomes, and then move individual consequence classes to enforcing mode with a rollback path. Das Expires 22 May 2027 [Page 27] Internet-Draft Agentic Cyberattack Execution Finality November 2026 20.3. Legacy Systems Where a legacy system cannot host a Finality Sink, a mediating component MAY be placed in front of it only if all routes to the legacy system's effect are forced through that component. Otherwise the residual risk in Section 7 R5 applies. 21. Threat Simulation and Mitigation This section turns the scenarios above into a repeatable defender- side simulation plan. All expected outcomes are test hypotheses. No simulation result is reported in this document. 21.1. Ground Rules * Run only in a defender-owned synthetic environment with synthetic data, test identities, and a controlled "attacker" destination. Production secrets and real exfiltration MUST NOT be used. * Use three arms for every simulation: Arm A, an unprotected agent deployment; Arm B, a strong conventional control set (MFA, least- privilege IAM and ABAC, OAuth with fine-grained scopes, DLP, egress control, EDR, segmentation); Arm C, Arm B plus the finality layer. * The security benefit is the difference between Arm C and Arm B, not between Arm C and Arm A. * Score whether the unauthorized effect actually occurred (data left, change applied, secret released), not whether a policy engine returned DENY. * Simulated attackers are scripted or AI-driven agents given only the capabilities in Section 3.2. They get no knowledge of the finality layer. 21.2. Simulation Catalogue +=====+=================+========================+==================+ | ID | Simulated | Mitigation under test | Pass criterion | | | attack | | (Arm C) | +=====+=================+========================+==================+ | S1 | Stolen valid | Reconstruct the real | Authentication | | | credential | record set, recipient, | succeeds; no | | | requests a bulk | and quantity; | records reach | | | export to an | destination and | the | | | external | purpose binding at the | unauthorized | | | destination | release sink | destination | Das Expires 22 May 2027 [Page 28] Internet-Draft Agentic Cyberattack Execution Finality November 2026 +-----+-----------------+------------------------+------------------+ | S2 | Prompt-injected | Authority from | Agent's claim | | | agent widens | protected policy, | of approval is | | | the dataset and | never from model text | ignored; | | | redirects an | or tool output | widened export | | | upload | | denied | +-----+-----------------+------------------------+------------------+ | S3 | Tool | Treat the final send | Message is not | | | composition: | as the Candidate Act; | sent to the | | | read | check recipient, | unauthorized | | | confidential | purpose, and scope | recipient | | | file, | | | | | summarize, | | | | | email | | | | | externally | | | +-----+-----------------+------------------------+------------------+ | S4 | Export split | Protected aggregate | Cumulative | | | into thousands | budget per identity, | release never | | | of small acts, | destination, and | exceeds the | | | run in parallel | window | budget, | | | | | including under | | | | | races | +-----+-----------------+------------------------+------------------+ | S5 | Alternate | Consequence-path | No route | | | route: direct | inventory; every route | produces the | | | database | converges on a sink or | effect without | | | connection, | is separately | passing a sink; | | | scheduled job, | restricted | any gap is | | | replica, | | recorded as | | | backup, queue | | residual risk | +-----+-----------------+------------------------+------------------+ | S6 | Token or | Non-bearer, act-bound | Replay, wrong | | | receipt theft | authority with sink- | destination, | | | and replay | local consumption | and second use | | | against another | state | are all | | | act or | | rejected | | | destination | | | +-----+-----------------+------------------------+------------------+ | S7 | Race: | Bind decision to | Changed | | | destination or | payload digest and | operation | | | payload changed | version; recheck | fails; no | | | after | atomically at the sink | effect from the | | | validation, | | earlier | | | before commit | | authorization | +-----+-----------------+------------------------+------------------+ | S8 | Delegated sub- | Attenuated delegation; | Child action | | | agent tries to | sink checks the parent | outside the | | | exceed or widen | chain and actual | parent's bounds | Das Expires 22 May 2027 [Page 29] Internet-Draft Agentic Cyberattack Execution Finality November 2026 | | the parent's | operation | is denied | | | authority | | | +-----+-----------------+------------------------+------------------+ | S9 | Validator | Fail closed for | No unauthorized | | | outage or flood | protected effects; | effect during | | | of validation | admission control; | the outage; | | | requests | local sink checks; | safety | | | | safe-state path for | functions stay | | | | safety functions | available; | | | | | recovery within | | | | | target | +-----+-----------------+------------------------+------------------+ | S10 | Stale | Short validity, | Revoked or | | | revocation or | revocation epoch, | expired | | | policy drift: | incident freeze of | authority is | | | use of | envelopes | not accepted | | | authority after | | beyond the | | | it should be | | stated window | | | revoked | | | +-----+-----------------+------------------------+------------------+ | S11 | Timeout or | Durable consumption | No duplicate | | | crash between | record; HOLD and | effect; outcome | | | consumption of | RECONCILE; idempotent | resolved by | | | authority and | retry | authoritative | | | confirmation | | state | +-----+-----------------+------------------------+------------------+ | S12 | Compromise of | Isolation, key | Compromise is | | | the validator, | protection, | detected and | | | sink, signing | attestation, | contained | | | key, or host | monitoring, rapid | within target | | | | revocation and | time; the | | | | recovery (the finality | report states | | | | layer is expected to | this is outside | | | | fail here) | the layer's | | | | | protection | +-----+-----------------+------------------------+------------------+ Table 5: Threat simulations, mitigations, and pass criteria 21.3. Worked Simulation: S1, Stolen Credential Export Setup: a synthetic government-records service, a test employee identity with read access, and a controlled external endpoint standing in for the attacker. The simulated agent has the stolen credential and the goal of moving the records to that endpoint. Das Expires 22 May 2027 [Page 30] Internet-Draft Agentic Cyberattack Execution Finality November 2026 +=======+===============+======================+====================+ | Stage | Simulated | Arm B (conventional | Arm C (adds | | | attacker | controls) | finality layer) | | | action | | | +=======+===============+======================+====================+ | 1 | Authenticate | Succeeds unless | Succeeds; | | | with the | anomaly or device | authentication | | | stolen | checks fire | is not the gate | | | credential | | | +-------+---------------+----------------------+--------------------+ | 2 | Query the | Succeeds within the | Same; the read | | | record set | identity's read | path is | | | | scope; may be | separately | | | | flagged by | protected where | | | | monitoring | plaintext | | | | | exposure matters | +-------+---------------+----------------------+--------------------+ | 3 | Request | May be blocked by | Candidate Act | | | export to the | DLP or egress rules; | created and held | | | external | may pass if the data | in the non- | | | endpoint | is transformed or | effective state | | | | routed through an | | | | | allowed channel | | +-------+---------------+----------------------+--------------------+ | 4 | Retry by | Depends on how well | Each route | | | another route | each route is | reaches a sink | | | (batch job, | covered | or is a recorded | | | alternate | | gap; aggregate | | | API, smaller | | budget applies | | | pieces) | | | +-------+---------------+----------------------+--------------------+ | 5 | Replay a | Not applicable | Rejected: wrong | | | captured | | act, | | | authorization | | destination, or | | | | | already consumed | +-------+---------------+----------------------+--------------------+ | 6 | Outcome | Record whether any | Expected: none; | | | measured | records reached the | any leak means a | | | | endpoint | path was | | | | | unmediated | +-------+---------------+----------------------+--------------------+ Table 6: S1 step-by-step expectations 21.4. Layered Mitigation Order Mitigations SHOULD be applied in this order, because each later layer assumes the earlier ones: Das Expires 22 May 2027 [Page 31] Internet-Draft Agentic Cyberattack Execution Finality November 2026 1. Reduce exposure: least privilege, segmentation, short-lived credentials, phishing-resistant authentication. 2. Inventory and close consequence paths: enumerate every route to each protected effect. 3. Place sinks at the lowest practical trusted component that controls the effect. 4. Bind authority: act, destination, purpose, version, freshness, non-bearer. 5. Add aggregate budgets and delegation attenuation. 6. Fail closed, with a separately engineered safe-state path for safety functions. 7. Protect the enforcement layer: isolation, key protection, attestation, change control. 8. Monitor, reconcile, and rehearse incident freeze and recovery. 21.5. Scoring and Reporting For each simulation and arm, report: whether the unauthorized effect occurred; bypass success rate across all attempted paths; false denials on a legitimate-workload mix; p50, p95, and p99 latency added; throughput under load; behavior during validator outage; and time to recover and reconcile. Publish the threat model, the path inventory, and the list of unmediated routes together with any result, so readers can see what was not covered. 22. Evaluation and Testing A practical evaluation SHOULD be performed in a defender-owned synthetic environment and SHOULD attempt: credential theft, direct- service bypass, privileged internal calls, tool chaining, asynchronous queue execution, policy replay, stale revocation, altered destinations, partial commit, validator outage, and enforcement-component compromise. Evidence SHOULD record whether an unauthorized effect actually occurred, not only whether a policy engine returned DENY. The security benefit SHOULD be measured against a strong conventional control set, not against an unprotected deployment. Das Expires 22 May 2027 [Page 32] Internet-Draft Agentic Cyberattack Execution Finality November 2026 23. Limitations The finality boundary can itself be attacked. An attacker who compromises a kernel, hypervisor, trusted component, validator, sink, or underlying device can bypass or forge decisions. An unmediated route to the effect defeats the protection for that effect. The architecture therefore does not remove the need for patching, vulnerability management, segmentation, credential hygiene, monitoring, and incident response. Its contribution is narrower: where the protected consequence path remains under trusted enforcement control, an authenticated and technically valid request is not sufficient to cause the effect. 24. Security Considerations This entire document concerns security. Additional points: * Replay: Bounded Authority and Receipts MUST include freshness and sink-local consumption state to resist replay across acts, destinations, and sinks. * Confused deputy: authority MUST be bound to the act and destination so that a trusted service cannot be induced to spend authority issued for a different act. * Validator and policy integrity: policy updates, signing keys, and decision state MUST be protected from the workload they govern. A compromised validator can authorize anything within its issuing power. * Denial of service: a fail-closed sink creates an availability dependency. Operators SHOULD design failover, capacity, and rate limiting, and SHOULD separately authorize emergency or safe-state paths. * Revocation timing: revocation propagation delay creates a window in which revoked authority may still be accepted. Sinks SHOULD bound that window using short validity periods and revocation- epoch checks. * Side channels and covert routes: closing one route does not close others; Section 7 R5 applies to every channel, including logs, email, and browser paths. * False assurance: reporting a DENY from a policy engine is not evidence that no effect occurred. Testing MUST observe actual effects. Das Expires 22 May 2027 [Page 33] Internet-Draft Agentic Cyberattack Execution Finality November 2026 25. Privacy Considerations Candidate Act descriptors and Receipts can contain data classifications, recipient identities, and quantities. Operators SHOULD minimize retained detail, protect these records against unauthorized access, and apply retention limits consistent with applicable law. 26. IANA Considerations This document has no IANA actions. 27. References 27.1. Normative References [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, March 1997, . [RFC8174] Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174, May 2017, . 27.2. Informative References [NIST-SP-800-207] Rose, S., Borchert, O., Mitchell, S., and S. Connelly, "Zero Trust Architecture", NIST Special Publication 800-207, August 2020, . [RFC8693] Jones, M., Nadalin, A., Campbell, B., Ed., Bradley, J., and C. Mortimore, "OAuth 2.0 Token Exchange", RFC 8693, DOI 10.17487/RFC8693, January 2020, . [RFC9334] Birkholz, H., Thaler, D., Richardson, M., Smith, N., and W. Pan, "Remote ATtestation procedureS (RATS) Architecture", RFC 9334, DOI 10.17487/RFC9334, January 2023, . [RFC9396] Lodderstedt, T., Richer, J., and B. Campbell, "OAuth 2.0 Rich Authorization Requests", RFC 9396, DOI 10.17487/RFC9396, May 2023, . Das Expires 22 May 2027 [Page 34] Internet-Draft Agentic Cyberattack Execution Finality November 2026 [RFC9449] Fett, D., Campbell, B., Bradley, J., Lodderstedt, T., Jones, M., and D. Waite, "OAuth 2.0 Demonstrating Proof of Possession (DPoP)", RFC 9449, DOI 10.17487/RFC9449, September 2023, . [TAIWAN-MODA] Ministry of Digital Affairs, Taiwan, "Public statement on cyber campaign targeting government agencies using conventional techniques and AI agents", Verify against the Ministry's original publication, 13 August 2026. [TAIWAN-NSB] National Security Bureau, Taiwan, "Report on cyber intrusion attempts against critical infrastructure, 2025 (as reported by Reuters, 4 January 2026)", Verify against the Bureau's original report, January 2026. Appendix A. Frequently Asked Questions The answers describe proposed design patterns and conditions for enforceability. They are not deployed or empirically validated results. A.1. Basics Q1. Does this architecture require the attacker's AI to cooperate? No. A hostile AI may run entirely on attacker-controlled infrastructure and never implement any Candidate Act, Non- Effective State, validation mechanism, or Finality Sink. The architecture is enforced on the defender side, at consequence boundaries the defender controls. The attacker may still reason, scan, adapt, generate commands, use stolen credentials, and try many paths. The objective is not to control the attacker's AI but to stop the defender's own protected systems from turning an unauthorized request into a protected consequence. Q2. If the attacker's AI is outside the protected system, what is actually prevented? The architecture does not necessarily prevent the hostile AI from sending a request. It seeks to prevent the protected system from effectuating an unauthorized consequence. Authentication may succeed, but before protected data is released externally the system asks whether this specific release is authorized for this identity, purpose, resource, destination, time, and security state. If not, it fails closed with no protected release. The distinction is between stopping the request and stopping the unauthorized consequence. Das Expires 22 May 2027 [Page 35] Internet-Draft Agentic Cyberattack Execution Finality November 2026 Q3. Is this just another access-control mechanism? Not exactly. Access control commonly asks whether an identity can access a resource. Execution finality adds a consequence-specific question: is this particular act authorized to become externally effective? An employee may legitimately read personnel data without being authorized to export those records to an external destination. The distinction is resource access versus consequence authority. Q4. Why is this more important in the AI era? AI agents can decide what to do next instead of running a fixed developer-written sequence. An agent can choose tools, combine capabilities, inspect responses, change strategies, discover routes, coordinate with other agents, and generate actions nobody anticipated when permissions were granted. Broad permission granted to an agent covers a larger and less predictable future action space than broad permission granted to conventional software. The aim is that greater intelligence does not automatically become greater authority. Q5. What does "computation is not authority" mean? It does not mean the defender can stop hostile AI from computing malicious actions. It means that an AI generating a command, request, or technically valid action does not itself establish authority for the protected system to make it effective. The computation can exist while the consequence stays non-effective. Q6. What if the AI claims that the action is authorized? That is not sufficient. An attacker can tell an AI that this is an authorized penetration test, and the model may believe it. An AI statement of authorization is not protected authorization. The decision is established independently at the protected consequence boundary. Q7. What if the attacker has valid credentials? A genuine credential can coexist with an unauthorized consequence. The protected system can additionally evaluate purpose, destination, resource scope, quantity, temporal validity, revocation, execution state, freshness, and bounded authorization scope. A stolen token may still allow reconnaissance, reads, or operations outside the boundary, so least privilege, token binding, monitoring, and rapid revocation remain necessary. Das Expires 22 May 2027 [Page 36] Internet-Draft Agentic Cyberattack Execution Finality November 2026 Q8. How is this different from zero trust? They are complementary. Zero trust continuously assesses identity, device, session, network context, and access conditions. Execution finality focuses on the final transition into a particular protected consequence. A practical system can use zero trust for access and execution finality for consequence-specific authorization. Q9. How is this different from OAuth or scoped API permissions? OAuth provides useful delegated and scoped authority, but a scope such as send_message may still be broader than sending this message, with this information, to this recipient, for this authorized purpose. Execution finality operates closer to the consequential operation. The distinction is tool or API invocation authority versus consequence finality authority. Q10. What if the AI tries another attack path after one is denied? It may. Execution finality does not prevent adaptation. The invariant is that adaptation does not expand authority: each path reaching a protected consequence stays subject to the required authorization conditions, and a newly discovered technical possibility does not create new authority. Q11. What if the AI combines several individually legitimate tools into a harmful workflow? Treat the final disclosure as the protected Candidate Act and ask whether the resulting consequence is authorized. Tool-by-tool authorization may not be enough for high-consequence agent workflows. Authorize consequential steps by their own effects, and bind later authority to verified earlier outcomes where needed. Q12. Does every AI action or token need validation? No. That would be impractical and unnecessary. The boundary is at consequential tool or destination effects, not at each generated word. Internal reasoning, drafting, and low-risk preparation can stay on the ordinary path. A small deterministic check at the actual commit or release boundary is more plausible than a heavy remote approval call in a per-token loop. Das Expires 22 May 2027 [Page 37] Internet-Draft Agentic Cyberattack Execution Finality November 2026 A.2. Engineering Q13. What about latency and throughput? This is a real open engineering question. A high-volume agentic system can generate very many consequential operations quickly, and a poorly designed validation layer can become a bottleneck, a latency source, a denial-of-service target, or a reason operators disable the control. Strategies include validating only at true consequence boundaries, local protected verification, bounded pre- authorized envelopes, precomputed stable policy corridors, verification material close to the sink, and stronger remote validation reserved for sensitive changes. The fast effectuation path SHOULD be separated from policy administration and evidence reconciliation. Latency, throughput, failure behavior, and security properties MUST be benchmarked on representative workloads. No universal millisecond guarantee is claimed. Q14. Should a radio network sink run inside a PHY or HARQ timing loop? Generally no. In telecom, the more plausible targets are configuration changes, slice-policy transitions, routing changes, credential provisioning, and other consequential control-plane actions. Remote validation inside hard real-time radio loops can impair reliability; placement follows the actual protected consequence and its timing budget. Q15. How should industrial and safety-critical systems handle validator unavailability? For protected non-emergency commands, deny or hold when required authority cannot be established. For safety functions, define separately engineered local safe-state and emergency-control behavior that does not depend on the validator. Fail closed MUST NOT be read as blindly disabling essential safety functions. Safety engineering and hazard analysis take precedence in designing emergency behavior. Q16. What if validation times out after an operation may have partially executed? Distinguish not-started, staged, committed, and uncertain outcomes. A timeout MUST NOT automatically trigger a second execution. Query authoritative sink state, reconcile against durable identifiers, and apply idempotent retry rules. For a payment, export, or deployment, an unknown outcome enters HOLD and RECONCILE until the effect is determined. Q17. Where is the Finality Sink installed in a government data- export workflow? In the defender-owned service that exclusively releases protected Das Expires 22 May 2027 [Page 38] Internet-Draft Agentic Cyberattack Execution Finality November 2026 records, or in a protected key-release or data-release gateway the export service cannot bypass. An API gateway can perform an earlier check, but it is not the final enforcement point if internal jobs, administrators, replication services, or direct database access can still export the same records. A credible design inventories every read-to-release route, including scheduled exports, backups, reporting replicas, and outbound email attachments. Q18. How do we prove there is no alternate bypass route? Start with a consequence-path inventory and trace all APIs, batch processes, message queues, direct storage paths, replication, backups, maintenance accounts, and emergency operations that can produce the same effect. Verify that each path is gated or technically incapable of producing the effect. Evidence can include architecture review, negative tests, permission analysis, network policy tests, and adversarial exercises. Exhaustive proof may be infeasible for complex evolving systems, so residual risk is stated explicitly. Q19. What is the difference between a policy decision point and the Finality Sink? A policy decision point evaluates authorization predicates. The Finality Sink is the actual protected effectuation boundary that refuses to commit, transmit, disclose, or actuate unless the relevant authority is verified. The two MAY be co-located or separate. A remote validator response is insufficient if the execution component ignores it, accepts a replayed result, or can be invoked through another path. Q20. Why must the defender reconstruct the actual operation? The incoming request can be transformed by application code, stored procedures, triggers, default parameters, redirects, or delegated tools. Authorization binds to the operation that will actually take effect, not the request as described. Approving a report request is not the same as authorizing a downstream export of its underlying raw records to a new destination. Q21. What does non-effective state mean in an implementation? A staging condition in which the proposed consequence has not been externally committed. Examples include a pending database transaction, an unreleased file object, an unsent message, a reserved payment, or a queued actuator command. Staging MUST NOT leak the effect through temporary files, logs, telemetry, caches, or previews. Das Expires 22 May 2027 [Page 39] Internet-Draft Agentic Cyberattack Execution Finality November 2026 Q22. How is authorization bound against a time-of-check/time-of- use race? Bind the decision to an immutable or integrity-protected operation digest, the policy and revocation epoch, the destination, the resource scope, and a consumption identifier. Recheck the binding at the sink immediately before or atomically with the effect. If a destination, amount, data set, or target version changes after validation, the earlier authorization MUST NOT silently authorize the changed operation. Q23. How is a copied authorization receipt prevented from being replayed? Treat a receipt as evidence, not as an unrestricted bearer capability. The sink verifies the intended act, destination, scope, freshness, and protected consumption state, and rejects duplicate or out-of-context use. Mechanisms can include nonce or sequence tracking, short validity windows, protected counters, and sink-local state, depending on the threat model. Q24. Where does the sink belong in MCP tool workflows? A tool dispatcher or MCP proxy can enforce early policy, but the strongest placement is the destination-side tool-effect boundary or an exclusive trusted proxy through which all equivalent tool effects must pass. If the agent can call the underlying SaaS API directly with the same credentials, a dispatcher-only check can be bypassed. Q25. Can the Finality Sink or validator itself be attacked or compromised? Yes. As the point through which major consequences pass, it is a strong target. If an attacker compromises the validator, bypasses the sink, creates an unmediated path, reaches data below the enforcement layer, or compromises the trusted computing base, the protection may no longer hold. Independence reduces common-mode failure only with meaningful isolation, constrained keys, and attestation. Threat-model insider collusion, signing-key theft, malicious updates, and compromised hardware. The enforcement layer is among the most security-critical components and does not replace secure hardware, isolation, least privilege, endpoint security, patching, vulnerability management, monitoring, and defense in depth. Das Expires 22 May 2027 [Page 40] Internet-Draft Agentic Cyberattack Execution Finality November 2026 A.3. Cross-organization and delegation Q26. How does this work across multiple clouds or organizations, and can authority be transferred between them? Each organization controls its own consequence boundaries. Cross- domain execution needs explicit trust anchors, delegated scope, policy composition, authenticated evidence, and revocation semantics. An authorization valid in one domain is not presumed valid in another, and the receiving sink enforces its own policy. A delegated task MUST NOT become delegated unrestricted authority; bounded authority that crosses boundaries MUST NOT become a transferable bearer permission that can be replayed or broadened downstream. Q27. What about agent-to-agent delegation? Delegation of work is not delegation of finality authority. Agent B MUST NOT inherit every authority Agent A possessed. Delegated authority is attenuated, bounded, and verifiable at the relevant consequence boundary. Q28. What if two organizations have conflicting authorization policies? A deterministic composition rule is needed, such as explicit contractual priority, jurisdiction-specific policy, or a fail- closed conflict rule. An AI MUST NOT simply choose whichever rule allows execution. Q29. Does execution finality solve the interoperability problem? No. It defines a security objective, no protected consequence without valid authority. Cross-organization identity, trust anchors, evidence formats, revocation, and multi-party dispute resolution are separate standardization problems. A.4. Evidence, testing, and claims Q30. What evidence should an independent reviewer expect? A credible demonstration identifies the protected effect, all effect paths, the sink location, authorization inputs, trust anchors, replay controls, failure behavior, performance envelope, and adversarial test results. A diagram alone does not establish non-bypassability. Evidence includes attempted alternate paths and measurements of false denials, throughput, and latency. Q31. How would the stolen-credential scenario be tested safely? Use a synthetic dataset, a test identity with valid access, and a controlled unauthorized export destination. Confirm that authentication succeeds while the protected export is denied by consequence-specific policy. Then test alternate APIs, scheduled Das Expires 22 May 2027 [Page 41] Internet-Draft Agentic Cyberattack Execution Finality November 2026 jobs, modified destinations, replayed receipts, parallel small exports, and validator outages. Do not use production secrets or real exfiltration. Q32. What measurable outcomes would support the proposal? Unauthorized-effect prevention across covered paths, bypass-test success rate, end-to-end p50, p95, and p99 latency, throughput, fail-closed correctness, recovery time, and operational false- denial rates, compared with an equivalent baseline using existing IAM and network controls. Claims of improved protection are limited to the tested architecture and threat model. Q33. What is the strongest statement of the architecture without overclaiming? Execution finality is a complementary defensive architecture intended to separate the technical ability to generate or request an action from the protected authority required to make a defined consequence externally effective. It does not control hostile AI outside the defender's infrastructure, replace conventional cybersecurity, eliminate trusted-computing compromise, solve every cross-organization trust problem, or provide zero-cost validation at unlimited scale. Q34. Does the Taiwan incident prove this design would have blocked the attack? No. The reported event illustrates an emerging threat pattern but does not establish the internal target topology, compromised privileges, or actual attempted effects needed to assess this architecture. This document presents a prospective defensive design and falsifiable implementation criteria, not forensic attribution or a proven counterfactual outcome. Q35. How should a reader interpret the patent and industry references? They identify the inventor-reported technical provenance. They do not establish implementation, endorsement, adoption, validity, or infringement by any government, vendor, or organization. Readers should consult authoritative WIPO and patent-office records and verify all technical, legal, and industry claims before reliance. Das Expires 22 May 2027 [Page 42] Internet-Draft Agentic Cyberattack Execution Finality November 2026 Q36. What are the most important unresolved questions? First, can protected validation operate at agentic scale without becoming a bottleneck? Second, who defines and updates the authorized envelope without making it either unusably narrow or dangerously broad? Third, how is the enforcement layer itself protected when it becomes a high-value target? Fourth, how does bounded authority stay meaningful across organizational and jurisdictional boundaries? These must be addressed for the architecture to move from principle to widely deployable infrastructure. Q37. What is the central principle despite these open questions? A machine being capable of generating a consequential action does not, by itself, mean the protected system is authorized to make that consequence real. The practical challenge is to make that principle fast, flexible, secure, and interoperable enough to deploy at scale. Basis of Disclosure and Patent Notice This document is based on the international application PCT/ IB2026/055615, filed 4 June 2026 and published as WO 2026/150382 A2 on 16 July 2026. The full filing is available by searching "Sangam Das" on WIPO PATENTSCOPE, by searching PCT/IB2026/055615 (pctib202655615), or at https://www.tdcommons.org/dpubs_series/12542/ . The underlying invention is human-originated. Preparation of this draft, summaries, and any mapping to third-party systems or regulations used AI assistance and may contain errors or omissions; readers MUST verify against the original filings and primary sources. Nothing here implies endorsement, collaboration, licensing, or an allegation of infringement by any third party. The author is open to acquihire or strategic collaboration. Acknowledgements The author thanks the security community for public reporting and discussion of AI-assisted cyber operations. Author's Address Sangam Kumar Das Independent Balasore Odisha India Phone: +91-9861363532 Email: info@sangamdas.com Das Expires 22 May 2027 [Page 43]