\n \n
\n \n \n \n \n \n \n\n\n\n\n Agent Audit Trail: A Standard Logging Format for Autonomous AI Systems\n
\n\n \n\n\n\n\n\n\n\n\n
\n \n \n
\n \n Agent Audit Trail: A Standard Logging Format for Autonomous AI Systems\n
\n draft-sharif-agent-audit-trail-00\n
\n- \n \n
- \n \n Status\n \n \n \n
- \n \n Email expansions\n \n \n \n
- \n \n History\n \n \n \n
\n This document is an Internet-Draft (I-D).\n Anyone may submit an I-D to the IETF.\n This I-D is not endorsed by the IETF and has no formal standing in the\n IETF standards process.\n
\n\n\n \n \n \n \n | Document | \nType | \n\n | \n \n\n\n\n\n\n\n\n Active Internet-Draft\n (individual)\n \n\n \n \n \n | \n
|---|---|---|---|
| \n | Author | \n\n \n | \n\n \n \n Raza Sharif \n \n \n \n | \n
| \n | Last updated | \n\n | \n 2026-03-29\n \n | \n
| \n | \n RFC stream\n | \n\n \n | \n\n \n (None)\n \n | \n
| \n | \n Intended RFC status\n | \n\n \n | \n\n \n \n (None)\n \n \n | \n
| \n | \n Formats\n | \n\n | \n\n \n \n \n\n \n | \n
| \n Stream\n | \n \n\n Stream state\n | \n\n | \n\n (No stream defined)\n | \n \n
| \n | \n Consensus boilerplate\n | \n\n \n | \n\n \n Unknown\n \n | \n
| \n | \n RFC Editor Note\n | \n\n \n | \n\n \n \n (None)\n \n \n | \n
| \n IESG\n | \n\n \n IESG state\n \n | \n\n \n | \n\n \n \n I-D Exists\n \n \n | \n
| \n | \n Telechat date\n | \n\n \n | \n\n \n \n (None)\n \n \n \n | \n
| \n | \n Responsible AD\n | \n\n \n | \n\n \n \n (None)\n \n \n | \n
| \n | \n Send notices to\n | \n\n \n | \n\n \n \n (None)\n \n \n | \n
\n
\n \n \n \n \n \n \n draft-sharif-agent-audit-trail-00\n \n
\n \n
\n Internet Engineering Task Force R. Sharif\nInternet-Draft CyberSecAI Ltd\nIntended status: Standards Track March 29, 2026\nExpires: September 29, 2026\n\n Agent Audit Trail: A Standard Logging Format for Autonomous AI\n Systems\n draft-sharif-agent-audit-trail-00\n\nAbstract\n\n This document specifies a standard logging format for autonomous\n AI agent systems. The Agent Audit Trail (AAT) defines a\n JSON-based record structure with mandatory fields for agent\n identity, action classification, outcome tracking, and trust\n level reporting. Records are linked via tamper-evident hash\n chaining using SHA-256 per RFC 8785, with optional ECDSA\n signatures for non-repudiation.\n\n The format addresses requirements from the EU AI Act\n (Regulation 2024/1689), which mandates automatic recording of\n events for high-risk AI systems effective August 2026. It also\n maps to SOC 2 Trust Services Criteria, ISO/IEC 42001, and\n PCI DSS v4.0.1 logging requirements.\n\n The design is transport-agnostic and supports export to JSONL,\n Syslog (RFC 5424), and CSV while preserving chain integrity.\n Privacy is addressed through input/output hashing and tombstone-\n based deletion compatible with GDPR Article 17.\n\nStatus of This Memo\n\n This Internet-Draft is submitted in full conformance with the\n provisions of BCP 78 and BCP 79.\n\n Internet-Drafts are working documents of the Internet Engineering\n Task Force (IETF). Note that other groups may also distribute\n working documents as Internet-Drafts. The list of current\n Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.\n\n Internet-Drafts are draft documents valid for a maximum of six\n months and may be updated, replaced, or obsoleted by other\n documents at any time. It is inappropriate to use Internet-Drafts\n as reference material or to cite them other than as "work in\n progress."\n\n This Internet-Draft will expire on September 29, 2026.\n\nCopyright Notice\n\n Copyright (c) 2026 IETF Trust and the persons identified as the\n document authors. All rights reserved.\n\n This document is subject to BCP 78 and the IETF Trust's Legal\n Provisions Relating to IETF Documents\n (https://trustee.ietf.org/license-info) in effect on the date of\n publication of this document. Please review these documents\n carefully, as they describe your rights and restrictions with\n respect to this document. Code Components extracted from this\n document must include Revised BSD License text as described in\n Section 4.e of the Trust Legal Provisions and are provided\n without warranty as described in the Revised BSD License.\n\nTable of Contents\n\n 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . 3\n 1.1. The Problem . . . . . . . . . . . . . . . . . . . . . . 3\n 1.2. Design Goals . . . . . . . . . . . . . . . . . . . . . 4\n 2. Terminology . . . . . . . . . . . . . . . . . . . . . . . . 4\n 3. Audit Record Format . . . . . . . . . . . . . . . . . . . . 5\n 3.1. Mandatory Fields . . . . . . . . . . . . . . . . . . . 5\n 3.2. Optional Fields . . . . . . . . . . . . . . . . . . . . 8\n 3.3. Field Constraints . . . . . . . . . . . . . . . . . . . 9\n 4. Tamper-Evident Chaining . . . . . . . . . . . . . . . . . . 10\n 4.1. Hash Computation . . . . . . . . . . . . . . . . . . . 10\n 4.2. Signature Envelope . . . . . . . . . . . . . . . . . . 11\n 4.3. Chain Verification . . . . . . . . . . . . . . . . . . 11\n 5. Action Type Definitions . . . . . . . . . . . . . . . . . . 12\n 5.1. tool_call . . . . . . . . . . . . . . . . . . . . . . . 12\n 5.2. tool_response . . . . . . . . . . . . . . . . . . . . . 13\n 5.3. decision . . . . . . . . . . . . . . . . . . . . . . . 13\n 5.4. delegation . . . . . . . . . . . . . . . . . . . . . . 13\n 5.5. escalation . . . . . . . . . . . . . . . . . . . . . . 14\n 5.6. error . . . . . . . . . . . . . . . . . . . . . . . . . 14\n 5.7. lifecycle . . . . . . . . . . . . . . . . . . . . . . . 14\n 6. Session Structure . . . . . . . . . . . . . . . . . . . . . 15\n 6.1. Genesis Record . . . . . . . . . . . . . . . . . . . . 15\n 6.2. Ordered Chain . . . . . . . . . . . . . . . . . . . . . 15\n 6.3. Session Close . . . . . . . . . . . . . . . . . . . . . 16\n 7. Retention Requirements . . . . . . . . . . . . . . . . . . 16\n 7.1. High-Risk Systems . . . . . . . . . . . . . . . . . . . 16\n 7.2. General-Purpose Systems . . . . . . . . . . . . . . . . 17\n 7.3. Tombstone Records . . . . . . . . . . . . . . . . . . . 17\n 8. Export Formats . . . . . . . . . . . . . . . . . . . . . . 17\n 8.1. JSONL (Primary) . . . . . . . . . . . . . . . . . . . . 18\n 8.2. Syslog (RFC 5424) . . . . . . . . . . . . . . . . . . . 18\n 8.3. CSV . . . . . . . . . . . . . . . . . . . . . . . . . . 18\n 9. Regulatory Mapping . . . . . . . . . . . . . . . . . . . . 19\n 9.1. EU AI Act . . . . . . . . . . . . . . . . . . . . . . . 19\n 9.2. SOC 2 . . . . . . . . . . . . . . . . . . . . . . . . . 20\n 9.3. ISO/IEC 42001 . . . . . . . . . . . . . . . . . . . . . 20\n 9.4. PCI DSS v4.0.1 . . . . . . . . . . . . . . . . . . . . 20\n 10. Privacy Considerations . . . . . . . . . . . . . . . . . . 21\n 10.1. Data Minimization . . . . . . . . . . . . . . . . . . . 21\n 10.2. Right to Erasure . . . . . . . . . . . . . . . . . . . 21\n 11. Security Considerations . . . . . . . . . . . . . . . . . . 22\n 11.1. Log Tampering . . . . . . . . . . . . . . . . . . . . . 22\n 11.2. Log Injection . . . . . . . . . . . . . . . . . . . . . 22\n 11.3. Timing Attacks . . . . . . . . . . . . . . . . . . . . 23\n 11.4. Chain Breaks . . . . . . . . . . . . . . . . . . . . . 23\n 12. IANA Considerations . . . . . . . . . . . . . . . . . . . . 23\n 12.1. Action Type Registry . . . . . . . . . . . . . . . . . 23\n 12.2. Outcome Registry . . . . . . . . . . . . . . . . . . . 24\n 13. References . . . . . . . . . . . . . . . . . . . . . . . . 24\n 13.1. Normative References . . . . . . . . . . . . . . . . . 24\n 13.2. Informative References . . . . . . . . . . . . . . . . 25\n Appendix A. Example Audit Trail . . . . . . . . . . . . . . . 26\n Appendix B. EU AI Act Compliance Checklist . . . . . . . . . . 31\n Appendix C. Implementation Notes . . . . . . . . . . . . . . . 32\n Author's Address . . . . . . . . . . . . . . . . . . . . . . . 34\n\n1. Introduction\n\n1.1. The Problem\n\n The EU Artificial Intelligence Act (Regulation 2024/1689) enters\n full application on 2 August 2026. Article 12 requires that\n high-risk AI systems "shall technically allow for the automatic\n recording of events ('logs') over the lifetime of the system."\n Article 12(2) further specifies that logging capabilities shall\n conform to recognized standards or common specifications.\n\n Despite this regulatory mandate, no standard exists for HOW\n autonomous AI agents should log their activities. Current\n approaches suffer from several deficiencies:\n\n o Proprietary formats that vary across vendors, making cross-\n system auditing impossible.\n\n o No tamper-evidence, allowing post-hoc modification of logs\n without detection.\n\n o Inconsistent action taxonomies that prevent meaningful\n comparison of agent behavior across implementations.\n\n o No linkage between agent identity and logged actions,\n making attribution unreliable.\n\n o No session structure, making it impossible to reconstruct\n the full sequence of an agent's autonomous decision chain.\n\n This document fills this gap by defining the Agent Audit Trail\n (AAT), a standard JSON-based logging format with tamper-evident\n chaining, a defined action taxonomy, and explicit regulatory\n mapping.\n\n1.2. Design Goals\n\n The AAT format is designed with the following goals:\n\n o Regulatory compliance: Direct mapping to EU AI Act\n Article 12 requirements and other frameworks.\n\n o Tamper evidence: Hash-chained records that make\n unauthorized modification detectable.\n\n o Interoperability: A single format usable across different\n agent frameworks, model providers, and orchestration\n systems.\n\n o Privacy by design: No raw personal data in records;\n cryptographic hashes of inputs and outputs instead.\n\n o Transport agnostic: Exportable to JSONL, Syslog, and CSV\n without losing chain integrity.\n\n o Incremental adoption: Mandatory fields are minimal;\n optional fields support progressive enhancement.\n\n2. Terminology\n\n The key words "MUST", "MUST NOT", "REQUIRED", "SHALL",\n "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT\n RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be\n interpreted as described in BCP 14 [RFC2119] [RFC8174] when,\n and only when, they appear in capitalized form, as shown here.\n\n Agent: An autonomous software system that uses one or more\n large language models to make decisions and take actions\n with limited or no human intervention per action.\n\n Agent Audit Trail (AAT): An ordered sequence of audit records\n produced by an agent during a session, linked by hash\n chaining.\n\n Audit Record: A single JSON object representing one logged\n event in an agent's operation.\n\n Session: A bounded sequence of agent operations that begins\n with a genesis record and ends with a session close record.\n\n Genesis Record: The first record in a session, which has no\n parent and establishes the chain root.\n\n Tombstone Record: A record that replaces a deleted record's\n content while preserving the hash chain.\n\n Trust Level: A classification from L0 (no verification) to\n L4 (full mutual authentication with revocation checking)\n as defined in [MCPS].\n\n Action Type: A controlled vocabulary value describing the\n category of agent activity being logged.\n\n Chain Hash: The SHA-256 digest of the previous record's\n canonical JSON representation, linking records into a\n tamper-evident sequence.\n\n3. Audit Record Format\n\n Each audit record is a JSON object. Fields are divided into\n mandatory (MUST be present in every record) and optional (MAY\n be present based on the action type and deployment context).\n\n3.1. Mandatory Fields\n\n record_id: String. REQUIRED. A UUID version 4 [RFC9562]\n uniquely identifying this record. Implementations MUST\n generate a fresh UUIDv4 for each record. Duplicate\n record_id values within a session indicate a processing\n error and MUST be flagged by validators.\n\n Example: "f47ac10b-58cc-4372-a567-0e02b2c3d479"\n\n timestamp: String. REQUIRED. The time at which the event\n occurred, formatted per RFC 3339 [RFC3339] with mandatory\n UTC offset. Implementations SHOULD use UTC (indicated by\n "Z" suffix). Millisecond precision is RECOMMENDED.\n Microsecond precision is OPTIONAL.\n\n Example: "2026-03-29T14:30:00.123Z"\n\n agent_id: String. REQUIRED. A URI [RFC3986] uniquely\n identifying the agent instance. This SHOULD be a\n persistent identifier that survives agent restarts. When\n used with MCPS [MCPS], this MUST match the agent_id in the\n Agent Passport.\n\n Example: "urn:agent:payment-bot.acme.example"\n\n agent_version: String. REQUIRED. The semantic version\n [SEMVER] of the agent software. This allows correlation\n of behavior changes with software updates.\n\n Example: "2.1.0"\n\n session_id: String. REQUIRED. A UUID version 4 identifying\n the current session. All records within a single session\n MUST share the same session_id.\n\n Example: "a1b2c3d4-e5f6-7890-abcd-ef1234567890"\n\n action_type: String. REQUIRED. One of the registered\n action type values defined in Section 5. The initial\n registry contains: "tool_call", "tool_response",\n "decision", "delegation", "escalation", "error",\n "lifecycle".\n\n action_detail: Object. REQUIRED. A JSON object containing\n action-type-specific fields as defined in Section 5. The\n structure of this object varies by action_type. Unknown\n fields within action_detail SHOULD be preserved by\n processors.\n\n outcome: String. REQUIRED. The result of the action. One\n of the registered outcome values: "success", "failure",\n "timeout", "denied", "escalated". See Section 12.2 for\n the outcome registry.\n\n o "success": The action completed as intended.\n\n o "failure": The action did not complete due to an error.\n\n o "timeout": The action exceeded its time budget.\n\n o "denied": The action was blocked by a policy or\n authorization check.\n\n o "escalated": The action was redirected to a human or\n higher-authority agent.\n\n trust_level: String. REQUIRED. The trust level at which\n the agent was operating when this action occurred. One of:\n "L0", "L1", "L2", "L3", "L4".\n\n o L0: No verification. The agent has no cryptographic\n identity.\n\n o L1: Self-signed identity. The agent possesses a key\n pair but no external attestation.\n\n o L2: Authority-signed identity. A Trust Authority has\n issued the agent's passport.\n\n o L3: Mutual authentication. Both parties have verified\n each other's identity.\n\n o L4: Full mutual authentication with revocation\n checking and continuous monitoring.\n\n parent_record_id: String or null. REQUIRED. The record_id\n of the immediately preceding record in the session chain.\n For genesis records (the first record in a session), this\n field MUST be null. For all subsequent records, this MUST\n contain the record_id of the previous record.\n\n prev_hash: String or null. REQUIRED. The SHA-256 hash of\n the canonical JSON representation (per RFC 8785 [RFC8785])\n of the previous record. For genesis records, this field\n MUST be null. The hash MUST be encoded as a lowercase\n hexadecimal string (64 characters).\n\n Example: "a7ffc6f8bf1ed76651c14756a061d662..."\n\n3.2. Optional Fields\n\n The following fields are OPTIONAL and MAY be included in any\n audit record:\n\n human_override: Object. Present when a human intervened in\n or overrode the agent's action. Contains:\n\n o "operator_id": String. An identifier for the human\n operator (SHOULD be a pseudonym or role, not a real\n name, for privacy).\n\n o "reason": String. Free-text explanation of the\n override.\n\n o "original_action": Object. The action the agent would\n have taken without intervention.\n\n risk_score: Number. A value between 0.0 and 1.0 indicating\n the agent's assessed risk of the action. 0.0 indicates\n minimal risk; 1.0 indicates maximum risk.\n\n model_id: String. The identifier of the language model used\n for the decision. Example: "gpt-4o-2025-03-01" or\n "claude-sonnet-4-20250514".\n\n input_hash: String. The SHA-256 hash of the input provided\n to the agent for this action, encoded as lowercase\n hexadecimal. Used instead of raw input to preserve\n privacy.\n\n output_hash: String. The SHA-256 hash of the output\n produced by the agent for this action, encoded as\n lowercase hexadecimal.\n\n latency_ms: Number. The wall-clock time in milliseconds\n from action initiation to completion.\n\n cost_estimate: Object. Estimated cost of the action:\n\n o "amount": Number. The monetary amount.\n\n o "currency": String. ISO 4217 currency code.\n\n o "breakdown": Object. Optional sub-costs (e.g.,\n "compute", "api_calls", "tokens").\n\n sanctions_check: Object. Result of sanctions screening:\n\n o "provider": String. The screening provider.\n\n o "checked_at": String. RFC 3339 timestamp of check.\n\n o "result": String. One of "clear", "match", "error".\n\n o "list_version": String. Version of the sanctions list.\n\n jurisdiction: String. ISO 3166-1 alpha-2 country code\n indicating the jurisdiction governing this action.\n\n signature: String. An ECDSA P-256 signature over the\n canonical JSON of this record (excluding the signature\n field itself), encoded as Base64url per RFC 4648\n Section 5. See Section 4.2.\n\n3.3. Field Constraints\n\n The following constraints apply to all audit records:\n\n o All string fields MUST be valid UTF-8.\n\n o The total size of a single audit record SHOULD NOT exceed\n 64 KB when serialized as JSON. Records exceeding 256 KB\n MUST be rejected by validators.\n\n o Timestamps MUST NOT be backdated. The timestamp of record\n N+1 MUST be greater than or equal to the timestamp of\n record N within the same session.\n\n o The action_detail object MUST contain at least one field\n relevant to the action_type.\n\n o Implementations MUST NOT add fields with names beginning\n with "aat_" to action_detail, as this prefix is reserved\n for future extensions of this specification.\n\n4. Tamper-Evident Chaining\n\n4.1. Hash Computation\n\n The prev_hash field creates a tamper-evident chain across all\n records in a session. The hash is computed as follows:\n\n 1. Take the complete JSON object of the previous record,\n INCLUDING all fields (mandatory and optional) that were\n present in the record as stored.\n\n 2. Serialize the JSON object using the JSON Canonicalization\n Scheme (JCS) defined in RFC 8785 [RFC8785]. JCS\n produces a deterministic byte sequence from any JSON\n value.\n\n 3. Compute the SHA-256 hash of the canonical byte sequence.\n\n 4. Encode the resulting 32-byte hash as a 64-character\n lowercase hexadecimal string.\n\n The formula is:\n\n prev_hash(N) = hex(SHA-256(JCS(record(N-1))))\n\n For the genesis record (N=0), prev_hash MUST be null.\n\n Implementations MUST use JCS (RFC 8785) for canonicalization.\n Alternative canonicalization schemes MUST NOT be used, as they\n would break chain verification across implementations.\n\n4.2. Signature Envelope\n\n When cryptographic non-repudiation is required, records MAY\n include an ECDSA P-256 signature. The signing procedure is:\n\n 1. Construct the complete audit record with all fields\n EXCEPT the "signature" field.\n\n 2. Serialize using JCS (RFC 8785).\n\n 3. Compute SHA-256 of the canonical bytes.\n\n 4. Sign the hash using ECDSA P-256 with the agent's private\n key, per FIPS 186-5 [FIPS186-5].\n\n 5. Encode the signature as Base64url (RFC 4648 Section 5)\n using IEEE P1363 fixed-length r||s encoding (64 bytes\n total: 32 bytes r, 32 bytes s).\n\n 6. Add the "signature" field to the record.\n\n When verifying, the verifier MUST remove the "signature" field\n before computing the hash for comparison.\n\n Note: When both prev_hash and signature are present, prev_hash\n is computed over the COMPLETE previous record INCLUDING its\n signature field. Only the current record's signature is\n excluded during signing of the current record.\n\n When used with MCPS [MCPS], the signing key SHOULD be the same\n key used in the agent's Agent Passport, providing a direct\n binding between audit records and cryptographic identity.\n\n4.3. Chain Verification\n\n To verify a session's audit trail integrity, a verifier MUST:\n\n 1. Confirm the first record has parent_record_id = null and\n prev_hash = null.\n\n 2. For each subsequent record N (where N > 0):\n\n a. Compute hex(SHA-256(JCS(record(N-1)))).\n\n b. Compare the computed hash with record(N).prev_hash.\n\n c. If the values differ, the chain is broken at\n record N and the trail MUST be flagged as tampered.\n\n 3. If signatures are present, verify each signature using\n the agent's public key.\n\n 4. Verify that timestamps are monotonically non-decreasing.\n\n 5. Verify that parent_record_id of record N equals the\n record_id of record N-1.\n\n A chain verification failure MUST be reported as a critical\n integrity error. Partial chain verification (e.g., verifying\n only the last K records) is NOT RECOMMENDED but MAY be used\n for performance reasons if the full chain has been previously\n verified.\n\n5. Action Type Definitions\n\n Each action type defines a specific structure for the\n action_detail object.\n\n5.1. tool_call\n\n Logged when the agent invokes an external tool or API.\n\n action_detail fields:\n\n o "tool_name": String. REQUIRED. The name of the tool\n being called.\n\n o "tool_server": String. OPTIONAL. URI of the MCP server\n or API endpoint providing the tool.\n\n o "parameters_hash": String. REQUIRED. SHA-256 hash of\n the serialized parameters sent to the tool.\n\n o "tool_version": String. OPTIONAL. Version of the tool\n definition.\n\n o "authorization": String. OPTIONAL. The authorization\n mechanism used (e.g., "bearer_token", "api_key",\n "mutual_tls").\n\n5.2. tool_response\n\n Logged when the agent receives a response from a tool.\n\n action_detail fields:\n\n o "tool_name": String. REQUIRED. The name of the tool\n that responded.\n\n o "response_hash": String. REQUIRED. SHA-256 hash of\n the response payload.\n\n o "response_size": Number. OPTIONAL. Size of the response\n in bytes.\n\n o "parent_call_id": String. REQUIRED. The record_id of\n the corresponding tool_call record.\n\n5.3. decision\n\n Logged when the agent makes an autonomous decision.\n\n action_detail fields:\n\n o "decision_type": String. REQUIRED. Category of decision\n (e.g., "route", "approve", "reject", "classify",\n "generate").\n\n o "reasoning_hash": String. OPTIONAL. SHA-256 hash of the\n agent's reasoning chain or chain-of-thought.\n\n o "confidence": Number. OPTIONAL. Confidence score between\n 0.0 and 1.0.\n\n o "alternatives_considered": Number. OPTIONAL. Count of\n alternative actions the agent evaluated.\n\n o "policy_ref": String. OPTIONAL. Identifier of the\n policy or rule that governed this decision.\n\n5.4. delegation\n\n Logged when the agent delegates work to another agent.\n\n action_detail fields:\n\n o "delegate_agent_id": String. REQUIRED. URI of the agent\n receiving the delegation.\n\n o "delegate_trust_level": String. REQUIRED. Trust level\n of the delegate agent.\n\n o "task_description_hash": String. REQUIRED. SHA-256 hash\n of the delegated task description.\n\n o "constraints": Array of String. OPTIONAL. Constraints\n imposed on the delegate.\n\n o "timeout_ms": Number. OPTIONAL. Maximum time allowed\n for the delegate to complete the task.\n\n5.5. escalation\n\n Logged when the agent escalates to a human operator or\n higher-authority system.\n\n action_detail fields:\n\n o "escalation_reason": String. REQUIRED. Why the agent\n escalated (e.g., "confidence_below_threshold",\n "policy_requires_human", "risk_score_exceeded",\n "error_recovery").\n\n o "escalation_target": String. REQUIRED. Identifier of\n the human or system receiving the escalation.\n\n o "context_hash": String. OPTIONAL. SHA-256 hash of the\n context provided to the escalation target.\n\n o "urgency": String. OPTIONAL. One of "low", "medium",\n "high", "critical".\n\n5.6. error\n\n Logged when the agent encounters an error condition.\n\n action_detail fields:\n\n o "error_code": String. REQUIRED. A machine-readable\n error code.\n\n o "error_message": String. REQUIRED. A human-readable\n error description.\n\n o "error_category": String. REQUIRED. One of "transport",\n "authentication", "authorization", "validation",\n "timeout", "internal", "external".\n\n o "recoverable": Boolean. REQUIRED. Whether the agent\n can continue operating after this error.\n\n o "stack_hash": String. OPTIONAL. SHA-256 hash of the\n stack trace, for debugging without exposing internals.\n\n5.7. lifecycle\n\n Logged for agent lifecycle events (start, stop, pause,\n configuration changes).\n\n action_detail fields:\n\n o "event": String. REQUIRED. One of "session_start",\n "session_end", "pause", "resume",\n "configuration_change", "key_rotation",\n "trust_level_change".\n\n o "previous_state": String. OPTIONAL. The state before\n this lifecycle event.\n\n o "new_state": String. OPTIONAL. The state after this\n lifecycle event.\n\n o "trigger": String. OPTIONAL. What caused the lifecycle\n event (e.g., "scheduled", "manual", "policy",\n "error_recovery").\n\n6. Session Structure\n\n6.1. Genesis Record\n\n Every session MUST begin with a genesis record. The genesis\n record has the following characteristics:\n\n o action_type MUST be "lifecycle".\n\n o action_detail.event MUST be "session_start".\n\n o parent_record_id MUST be null.\n\n o prev_hash MUST be null.\n\n o The action_detail SHOULD include the agent's configuration\n hash, enabled tools list, and operating parameters to\n establish a baseline for the session.\n\n Example genesis action_detail:\n\n {\n "event": "session_start",\n "new_state": "active",\n "trigger": "scheduled",\n "config_hash": "b5bb9d8014a0f9b1d6...",\n "enabled_tools": [\n "payment_transfer",\n "sanctions_check",\n "balance_query"\n ]\n }\n\n6.2. Ordered Chain\n\n After the genesis record, all records MUST form a strictly\n ordered chain:\n\n o Each record's parent_record_id MUST equal the previous\n record's record_id.\n\n o Each record's prev_hash MUST equal\n hex(SHA-256(JCS(previous_record))).\n\n o Timestamps MUST be monotonically non-decreasing.\n\n o No gaps in the chain are permitted. If a record cannot\n be produced (e.g., due to a crash), a recovery record\n with action_type "error" MUST be inserted to document the\n gap when the agent resumes.\n\n Branching (multiple records claiming the same parent) is NOT\n permitted within a single session. If an agent forks into\n parallel execution paths, each path MUST use a separate\n session_id and the delegation record in the parent session\n MUST reference the child session_id.\n\n6.3. Session Close\n\n Every session SHOULD end with a close record. The close\n record has the following characteristics:\n\n o action_type MUST be "lifecycle".\n\n o action_detail.event MUST be "session_end".\n\n o action_detail MUST include a "session_hash" field\n containing the SHA-256 hash of the concatenation of all\n record hashes in the session, in order:\n\n session_hash = hex(SHA-256(\n prev_hash(1) || prev_hash(2) || ... || prev_hash(N)\n ))\n\n where N is the close record itself and prev_hash values\n are the raw 32-byte digests (not hex-encoded) prior to\n concatenation.\n\n o action_detail SHOULD include "record_count" (integer)\n and "duration_ms" (number) summarizing the session.\n\n If an agent terminates abnormally without producing a close\n record, the session is considered "orphaned." Monitoring\n systems SHOULD detect orphaned sessions and produce a\n synthetic close record with outcome "failure" and\n action_detail.trigger "crash_recovery".\n\n7. Retention Requirements\n\n7.1. High-Risk Systems\n\n For AI systems classified as high-risk under the EU AI Act\n (Annex III), audit trail records SHOULD be retained for a\n minimum of 12 months from the session close timestamp.\n\n This aligns with Article 12(1) which states that logging\n capabilities shall be such that logs are kept for a period\n appropriate to the intended purpose of the high-risk AI\n system, of at least six months unless provided otherwise in\n applicable Union or national law.\n\n The 12-month RECOMMENDATION in this specification exceeds the\n minimum 6-month requirement to account for audit cycles and\n incident investigation timelines.\n\n7.2. General-Purpose Systems\n\n For AI systems not classified as high-risk, audit trail\n records SHOULD be retained for a minimum of 6 months from\n the session close timestamp.\n\n Deployments subject to financial regulations (e.g., PCI DSS,\n SOC 2) MAY require longer retention periods as specified by\n those frameworks.\n\n7.3. Tombstone Records\n\n When individual records must be deleted (e.g., pursuant to\n GDPR Article 17 right to erasure), the record MUST be\n replaced with a tombstone record that preserves chain\n integrity. A tombstone record:\n\n o Retains the original record_id, timestamp,\n parent_record_id, and prev_hash.\n\n o Sets action_type to "lifecycle".\n\n o Sets action_detail to:\n\n {\n "event": "record_deleted",\n "deletion_reason": "gdpr_art17",\n "deleted_at": "2026-06-15T10:00:00Z",\n "original_action_type": "tool_call"\n }\n\n o Sets outcome to "success".\n\n o Retains the signature field if originally present.\n\n The original record's content is destroyed. Because the\n tombstone preserves the original record_id and prev_hash,\n subsequent records in the chain remain verifiable. However,\n the prev_hash of the NEXT record will no longer match the\n tombstone (since the content changed). To handle this,\n implementations MUST also store a "tombstone_hash" field in\n the tombstone record containing the original record's hash,\n allowing validators to accept the chain break.\n\n8. Export Formats\n\n Implementations MUST support at least one export format.\n JSONL is the RECOMMENDED primary format.\n\n8.1. JSONL (Primary)\n\n The primary export format is JSON Lines (JSONL), where each\n line contains exactly one complete audit record serialized\n as JSON. Lines are separated by a single newline character\n (U+000A).\n\n o Each line MUST be a valid JSON object.\n\n o The order of lines MUST match the chain order\n (genesis first, close last).\n\n o The file SHOULD use UTF-8 encoding without a byte order\n mark (BOM).\n\n o The file extension SHOULD be ".jsonl".\n\n8.2. Syslog (RFC 5424)\n\n For integration with existing logging infrastructure, audit\n records MAY be exported as Syslog messages per RFC 5424\n [RFC5424]. The mapping is:\n\n o FACILITY: local0 (16).\n\n o SEVERITY: based on outcome -- success=6 (Informational),\n failure=3 (Error), timeout=4 (Warning),\n denied=5 (Notice), escalated=5 (Notice).\n\n o APP-NAME: the agent_id (truncated to 48 characters).\n\n o MSGID: the action_type.\n\n o STRUCTURED-DATA: SD-ID "aat@IANA-PEN" containing\n record_id, session_id, trust_level, prev_hash.\n\n o MSG: JSON serialization of the full audit record.\n\n The prev_hash and chain integrity MUST be preserved in the\n structured data to enable reconstruction of the chain from\n Syslog archives.\n\n8.3. CSV\n\n For human review and spreadsheet analysis, audit records MAY\n be exported as CSV per RFC 4180 [RFC4180]. The mapping is:\n\n o Header row: record_id, timestamp, agent_id,\n agent_version, session_id, action_type, outcome,\n trust_level, parent_record_id, prev_hash, action_detail.\n\n o The action_detail column contains the JSON serialization\n of the action_detail object.\n\n o CSV export is inherently lossy for optional fields.\n Implementations SHOULD document which optional fields\n are included.\n\n CSV exports MUST NOT be used as the authoritative record.\n The JSONL format MUST be retained as the source of truth.\n\n9. Regulatory Mapping\n\n9.1. EU AI Act\n\n The following table maps AAT features to EU AI Act articles:\n\n Article 12 (Record-Keeping):\n AAT provides automatic recording via the mandatory audit\n record format (Section 3). Hash chaining (Section 4)\n ensures records "allow the tracing back of the AI system's\n operation." The session structure (Section 6) provides the\n "period of each use" required by Art 12(1)(c).\n\n Article 13 (Transparency):\n The action_type taxonomy (Section 5) and decision records\n (Section 5.3) provide interpretability of agent behavior.\n The model_id field documents which model was used.\n The human_override field documents human interventions.\n\n Article 14 (Human Oversight):\n The escalation action type (Section 5.5) documents when\n and why agents escalated to humans. The human_override\n optional field (Section 3.2) captures human interventions.\n Trust levels document the degree of autonomous operation.\n\n Article 72 (Reporting):\n The export formats (Section 8) enable provision of logs\n to national competent authorities. The session_hash in\n session close records (Section 6.3) provides a verifiable\n summary for regulatory reporting.\n\n9.2. SOC 2\n\n SOC 2 Trust Services Criteria relevant to AAT:\n\n o CC6.1 (Logical Access): trust_level and authorization\n fields document access controls.\n\n o CC7.2 (System Monitoring): Continuous audit trail with\n tamper-evident chaining satisfies monitoring requirements.\n\n o CC8.1 (Change Management): lifecycle action type records\n document configuration changes.\n\n9.3. ISO/IEC 42001\n\n ISO/IEC 42001 (AI Management System) clauses addressed:\n\n o Clause 6.1.2 (AI Risk Assessment): risk_score and\n decision records support risk documentation.\n\n o Clause 8.4 (AI System Operation): Full session audit\n trails document operational behavior.\n\n o Clause 9.1 (Monitoring): Continuous logging with chain\n verification supports monitoring requirements.\n\n9.4. PCI DSS v4.0.1\n\n PCI DSS v4.0.1 requirements addressed by AAT:\n\n o Requirement 10.2: AAT provides audit logs for all agent\n actions including tool calls, decisions, and errors.\n\n o Requirement 10.3: Record fields (timestamp, agent_id,\n action_type, outcome) map directly to required audit\n trail entries.\n\n o Requirement 10.5: Hash chaining and optional signatures\n protect audit trail integrity.\n\n o Requirement 10.7: Retention requirements (Section 7)\n align with PCI DSS retention periods.\n\n10. Privacy Considerations\n\n10.1. Data Minimization\n\n AAT is designed with privacy by default:\n\n o Raw input and output data MUST NOT be stored in audit\n records. Implementations MUST use the input_hash and\n output_hash fields instead.\n\n o The human_override.operator_id SHOULD be a pseudonymous\n identifier or role name, not a natural person's name.\n\n o The reasoning_hash field in decision records stores a hash\n of the reasoning chain, not the reasoning itself.\n\n o Tool parameters are recorded via parameters_hash, not in\n cleartext.\n\n o Sanctions check results record only "clear", "match", or\n "error" -- not the details of what was screened.\n\n Implementations that need to retain raw data for debugging\n MUST store it in a separate system with appropriate access\n controls, linked to the audit trail via record_id.\n\n10.2. Right to Erasure\n\n To support GDPR Article 17 (right to erasure) and similar\n regulations, AAT uses tombstone records (Section 7.3) rather\n than record deletion. This approach:\n\n o Removes all personal data from the record.\n\n o Preserves chain integrity for regulatory compliance.\n\n o Documents the fact and reason for deletion.\n\n o Is compatible with the EU AI Act's record-keeping\n requirements, which do not require retention of personal\n data but do require retention of operational logs.\n\n Data controllers MUST implement a process to identify which\n audit records contain personal data (even in hashed form) and\n respond to erasure requests by creating tombstone records\n within 30 days.\n\n11. Security Considerations\n\n11.1. Log Tampering\n\n The primary threat to audit trails is unauthorized\n modification. AAT mitigates this through:\n\n o Hash chaining: Any modification to a record invalidates\n all subsequent prev_hash values, making tampering\n detectable.\n\n o Optional signatures: ECDSA P-256 signatures provide\n non-repudiation and prevent even the log storage system\n from undetectably modifying records.\n\n o Session hashes: The session_hash in close records provides\n a single value that can be stored externally (e.g., on a\n blockchain or with a timestamp authority) to anchor the\n entire session.\n\n Implementations SHOULD store session_hash values in a\n separate, append-only system to provide an independent\n verification point.\n\n11.2. Log Injection\n\n Attackers may attempt to inject false audit records into the\n trail. Mitigations include:\n\n o Signature verification: When signatures are present, only\n records signed by the agent's key are valid.\n\n o Chain continuity: Injected records would break the hash\n chain unless the attacker can also modify all subsequent\n records.\n\n o Timestamp monotonicity: Injected records with out-of-\n order timestamps are detectable.\n\n o Record size limits: The 256 KB maximum prevents denial-of-\n service through oversized records.\n\n Implementations MUST validate all records against the schema\n before accepting them into the audit trail.\n\n11.3. Timing Attacks\n\n Audit record timestamps may be manipulated if the agent\n controls its own clock. Mitigations include:\n\n o Using NTP-synchronized clocks with drift monitoring.\n\n o Cross-referencing timestamps with external systems (e.g.,\n tool server response timestamps).\n\n o Flagging sessions where timestamps show suspicious\n patterns (e.g., large jumps, regression).\n\n Implementations SHOULD monitor for timestamp anomalies and\n flag them for human review.\n\n11.4. Chain Breaks\n\n Chain breaks can occur due to:\n\n o System crashes during record writing.\n\n o Storage corruption.\n\n o Intentional tampering.\n\n When a chain break is detected, implementations MUST:\n\n 1. Flag the break with a severity of "critical".\n\n 2. Record the break location (which record pair failed\n verification).\n\n 3. Preserve both the broken chain and any recovered data.\n\n 4. If the break is due to crash recovery, insert an error\n record documenting the gap.\n\n Chain breaks in high-risk systems MUST trigger an alert to\n the system operator within 1 hour.\n\n12. IANA Considerations\n\n12.1. Action Type Registry\n\n This document requests IANA to create the "Agent Audit Trail\n Action Types" registry. The registration policy is\n "Specification Required" per RFC 8126 [RFC8126].\n\n Initial registry contents:\n\n +----------------+---------------------------+-----------+\n | Value | Description | Reference |\n +----------------+---------------------------+-----------+\n | tool_call | Agent invokes a tool | Sec 5.1 |\n | tool_response | Agent receives tool reply | Sec 5.2 |\n | decision | Agent makes a decision | Sec 5.3 |\n | delegation | Agent delegates to agent | Sec 5.4 |\n | escalation | Agent escalates to human | Sec 5.5 |\n | error | Agent encounters error | Sec 5.6 |\n | lifecycle | Agent lifecycle event | Sec 5.7 |\n +----------------+---------------------------+-----------+\n\n New entries MUST include a value (lowercase ASCII string,\n max 32 characters), description, and reference to a published\n specification.\n\n12.2. Outcome Registry\n\n This document requests IANA to create the "Agent Audit Trail\n Outcomes" registry. The registration policy is "Specification\n Required" per RFC 8126.\n\n Initial registry contents:\n\n +------------+-------------------------------+-----------+\n | Value | Description | Reference |\n +------------+-------------------------------+-----------+\n | success | Action completed as intended | Sec 3.1 |\n | failure | Action failed due to error | Sec 3.1 |\n | timeout | Action exceeded time budget | Sec 3.1 |\n | denied | Action blocked by policy | Sec 3.1 |\n | escalated | Action redirected to human | Sec 3.1 |\n +------------+-------------------------------+-----------+\n\n13. References\n\n13.1. Normative References\n\n [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate\n Requirement Levels", BCP 14, RFC 2119,\n DOI 10.17487/RFC2119, March 1997,\n <https://www.rfc-editor.org/info/rfc2119>.\n\n [RFC3339] Klyne, G. and C. Newman, "Date and Time on the\n Internet: Timestamps", RFC 3339,\n DOI 10.17487/RFC3339, July 2002,\n <https://www.rfc-editor.org/info/rfc3339>.\n\n [RFC3986] Berners-Lee, T., Fielding, R., and L. Masinter,\n "Uniform Resource Identifier (URI): Generic Syntax",\n STD 66, RFC 3986, DOI 10.17487/RFC3986,\n January 2005,\n <https://www.rfc-editor.org/info/rfc3986>.\n\n [RFC8174] Leiba, B., "Ambiguity of Uppercase vs Lowercase in\n RFC 2119 Key Words", BCP 14, RFC 8174,\n DOI 10.17487/RFC8174, May 2017,\n <https://www.rfc-editor.org/info/rfc8174>.\n\n [RFC8785] Rundgren, A., Jordan, B., and S. Erdtman, "JSON\n Canonicalization Scheme (JCS)", RFC 8785,\n DOI 10.17487/RFC8785, June 2020,\n <https://www.rfc-editor.org/info/rfc8785>.\n\n [RFC9562] Davis, K., Peabody, B., and P. Leach, "Universally\n Unique IDentifiers (UUIDs)", RFC 9562,\n DOI 10.17487/RFC9562, May 2024,\n <https://www.rfc-editor.org/info/rfc9562>.\n\n [FIPS186-5]\n National Institute of Standards and Technology,\n "Digital Signature Standard (DSS)", FIPS PUB 186-5,\n DOI 10.6028/NIST.FIPS.186-5, February 2023.\n\n [RFC8126] Cotton, M., Leiba, B., and T. Narten, "Guidelines\n for Writing an IANA Considerations Section in RFCs",\n BCP 26, RFC 8126, DOI 10.17487/RFC8126, June 2017,\n <https://www.rfc-editor.org/info/rfc8126>.\n\n13.2. Informative References\n\n [MCPS] Sharif, R., "MCPS: Cryptographic Security Layer for\n the Model Context Protocol",\n draft-sharif-mcps-secure-mcp-02, March 2026.\n\n [EU-AI-ACT]\n European Parliament and Council, "Regulation (EU)\n 2024/1689 laying down harmonised rules on artificial\n intelligence (Artificial Intelligence Act)",\n Official Journal of the European Union, L series,\n 2024/1689, August 2024.\n\n [RFC5424] Gerhards, R., "The Syslog Protocol", RFC 5424,\n DOI 10.17487/RFC5424, March 2009,\n <https://www.rfc-editor.org/info/rfc5424>.\n\n [RFC4180] Shafranovich, Y., "Common Format and MIME Type for\n Comma-Separated Values (CSV) Files", RFC 4180,\n DOI 10.17487/RFC4180, October 2005,\n <https://www.rfc-editor.org/info/rfc4180>.\n\n [RFC4648] Josefsson, S., "The Base16, Base32, and Base64 Data\n Encodings", RFC 4648, DOI 10.17487/RFC4648,\n October 2006,\n <https://www.rfc-editor.org/info/rfc4648>.\n\n [ISO42001] International Organization for Standardization,\n "Information technology -- Artificial intelligence --\n Management system", ISO/IEC 42001:2023, December\n 2023.\n\n [PCI-DSS] PCI Security Standards Council, "Payment Card\n Industry Data Security Standard Version 4.0.1",\n June 2024.\n\n [SOC2] American Institute of Certified Public Accountants,\n "SOC 2 -- SOC for Service Organizations: Trust\n Services Criteria", 2017.\n\n [SEMVER] Preston-Werner, T., "Semantic Versioning 2.0.0",\n <https://semver.org/>.\n\nAppendix A. Example Audit Trail\n\n The following example shows a complete audit trail for a\n payment agent session that processes a GBP 500 transfer.\n The session demonstrates tool calls, decisions, sanctions\n screening, and successful completion. prev_hash values are\n truncated for readability (shown as first 16 hex characters).\n\n Record 1: Genesis (session start)\n\n {\n "record_id": "a1000000-0000-4000-8000-000000000001",\n "timestamp": "2026-03-29T14:00:00.000Z",\n "agent_id": "urn:agent:payment-bot.acme.example",\n "agent_version": "2.1.0",\n "session_id": "sess-29mar-0001-4000-8000-abcdef123456",\n "action_type": "lifecycle",\n "action_detail": {\n "event": "session_start",\n "new_state": "active",\n "trigger": "api_request",\n "config_hash": "b5bb9d8014a0f9b1...",\n "enabled_tools": [\n "payment_transfer",\n "sanctions_check",\n "balance_query"\n ]\n },\n "outcome": "success",\n "trust_level": "L2",\n "parent_record_id": null,\n "prev_hash": null\n }\n\n Record 2: Sanctions screening tool call\n\n {\n "record_id": "a1000000-0000-4000-8000-000000000002",\n "timestamp": "2026-03-29T14:00:00.150Z",\n "agent_id": "urn:agent:payment-bot.acme.example",\n "agent_version": "2.1.0",\n "session_id": "sess-29mar-0001-4000-8000-abcdef123456",\n "action_type": "tool_call",\n "action_detail": {\n "tool_name": "sanctions_check",\n "tool_server": "https://screening.acme.example/v2",\n "parameters_hash": "e3b0c44298fc1c14...",\n "authorization": "mutual_tls"\n },\n "outcome": "success",\n "trust_level": "L2",\n "parent_record_id":\n "a1000000-0000-4000-8000-000000000001",\n "prev_hash": "7d865e959b2466918a...",\n "input_hash": "9f86d081884c7d659a...",\n "latency_ms": 145\n }\n\n Record 3: Sanctions screening response\n\n {\n "record_id": "a1000000-0000-4000-8000-000000000003",\n "timestamp": "2026-03-29T14:00:00.295Z",\n "agent_id": "urn:agent:payment-bot.acme.example",\n "agent_version": "2.1.0",\n "session_id": "sess-29mar-0001-4000-8000-abcdef123456",\n "action_type": "tool_response",\n "action_detail": {\n "tool_name": "sanctions_check",\n "response_hash": "2cf24dba5fb0a301...",\n "response_size": 256,\n "parent_call_id":\n "a1000000-0000-4000-8000-000000000002"\n },\n "outcome": "success",\n "trust_level": "L2",\n "parent_record_id":\n "a1000000-0000-4000-8000-000000000002",\n "prev_hash": "4e07408562bedb8b6...",\n "sanctions_check": {\n "provider": "acme_screening",\n "checked_at": "2026-03-29T14:00:00.290Z",\n "result": "clear",\n "list_version": "2026-03-29"\n }\n }\n\n Record 4: Decision to approve transfer\n\n {\n "record_id": "a1000000-0000-4000-8000-000000000004",\n "timestamp": "2026-03-29T14:00:00.310Z",\n "agent_id": "urn:agent:payment-bot.acme.example",\n "agent_version": "2.1.0",\n "session_id": "sess-29mar-0001-4000-8000-abcdef123456",\n "action_type": "decision",\n "action_detail": {\n "decision_type": "approve",\n "reasoning_hash": "6b86b273ff34fce1...",\n "confidence": 0.97,\n "alternatives_considered": 2,\n "policy_ref": "payment-policy-v3.2"\n },\n "outcome": "success",\n "trust_level": "L2",\n "parent_record_id":\n "a1000000-0000-4000-8000-000000000003",\n "prev_hash": "ef2d127de37b942ba...",\n "risk_score": 0.12,\n "model_id": "claude-sonnet-4-20250514",\n "cost_estimate": {\n "amount": 500.00,\n "currency": "GBP"\n }\n }\n\n Record 5: Payment execution tool call\n\n {\n "record_id": "a1000000-0000-4000-8000-000000000005",\n "timestamp": "2026-03-29T14:00:00.320Z",\n "agent_id": "urn:agent:payment-bot.acme.example",\n "agent_version": "2.1.0",\n "session_id": "sess-29mar-0001-4000-8000-abcdef123456",\n "action_type": "tool_call",\n "action_detail": {\n "tool_name": "payment_transfer",\n "tool_server": "https://payments.acme.example/v1",\n "parameters_hash": "d4735e3a265e16ee...",\n "authorization": "bearer_token"\n },\n "outcome": "success",\n "trust_level": "L2",\n "parent_record_id":\n "a1000000-0000-4000-8000-000000000004",\n "prev_hash": "e7f6c011776e8db7c...",\n "latency_ms": 890,\n "jurisdiction": "GB"\n }\n\n Record 6: Session close\n\n {\n "record_id": "a1000000-0000-4000-8000-000000000006",\n "timestamp": "2026-03-29T14:00:01.210Z",\n "agent_id": "urn:agent:payment-bot.acme.example",\n "agent_version": "2.1.0",\n "session_id": "sess-29mar-0001-4000-8000-abcdef123456",\n "action_type": "lifecycle",\n "action_detail": {\n "event": "session_end",\n "previous_state": "active",\n "new_state": "closed",\n "trigger": "task_complete",\n "session_hash": "9c22ff5f21f0b81b...",\n "record_count": 6,\n "duration_ms": 1210\n },\n "outcome": "success",\n "trust_level": "L2",\n "parent_record_id":\n "a1000000-0000-4000-8000-000000000005",\n "prev_hash": "a3a2e67ad1b8d57e2..."\n }\n\nAppendix B. EU AI Act Compliance Checklist\n\n The following table maps EU AI Act Article 12 sub-requirements\n to specific AAT features:\n\n +----------------------------+-------------------------------+\n | Art 12 Requirement | AAT Feature |\n +----------------------------+-------------------------------+\n | 12(1) Automatic recording | Mandatory audit record format |\n | | (Section 3) |\n +----------------------------+-------------------------------+\n | 12(1)(a) Recording of | session_id + ordered chain |\n | period of each use | (Section 6) |\n +----------------------------+-------------------------------+\n | 12(1)(b) Reference | agent_id (URI) + agent_version|\n | database against which | + model_id (Section 3) |\n | input data has been | |\n | checked | |\n +----------------------------+-------------------------------+\n | 12(1)(c) Input data for | input_hash field |\n | which search has led to | (Section 3.2) |\n | a match | |\n +----------------------------+-------------------------------+\n | 12(1)(d) Identification | human_override field with |\n | of natural persons | pseudonymous operator_id |\n | involved in verification | (Section 3.2) |\n +----------------------------+-------------------------------+\n | 12(2) Conform to | This specification |\n | recognised standards | |\n +----------------------------+-------------------------------+\n | 12(3) Appropriate to | Retention requirements |\n | intended purpose, at | (Section 7): 12 months for |\n | least 6 months | high-risk, 6 months general |\n +----------------------------+-------------------------------+\n | 12(4) Providers of | Export formats (Section 8) |\n | high-risk AI systems | enable log provision to |\n | that are credit | financial authorities |\n | institutions | |\n +----------------------------+-------------------------------+\n | Art 13 Transparency | action_type taxonomy + |\n | | decision records (Sec 5.3) |\n +----------------------------+-------------------------------+\n | Art 14 Human oversight | escalation type (Sec 5.5) + |\n | | human_override (Sec 3.2) |\n +----------------------------+-------------------------------+\n\nAppendix C. Implementation Notes\n\nC.1. Performance Considerations\n\n Hash computation adds overhead to each record. Benchmarks on\n commodity hardware show:\n\n o JCS canonicalization: ~0.1 ms per record (typical size).\n\n o SHA-256 hash: ~0.01 ms per record.\n\n o ECDSA P-256 signing: ~1-2 ms per record.\n\n o Total overhead with signing: ~2 ms per record.\n\n For high-throughput agents (>1000 actions/second),\n implementations MAY batch records and compute hashes\n asynchronously, provided the chain order is preserved. The\n timestamp MUST reflect the actual event time, not the time\n the hash was computed.\n\nC.2. Storage Estimates\n\n A typical audit record (mandatory fields only) is\n approximately 500-800 bytes when serialized as JSON. With\n optional fields, records range from 800-2000 bytes.\n\n For a high-activity agent producing 10,000 records per day:\n\n o Daily storage: ~10-20 MB (JSONL).\n\n o Monthly storage: ~300-600 MB.\n\n o 12-month retention: ~3.6-7.2 GB.\n\n Implementations SHOULD apply compression (e.g., gzip) to\n archived sessions. Typical compression ratios for JSON\n audit data are 5:1 to 10:1.\n\nC.3. Clock Synchronization\n\n Accurate timestamps are critical for audit trail integrity.\n Implementations MUST:\n\n o Use NTP or PTP for clock synchronization.\n\n o Monitor clock drift and alert if drift exceeds 100 ms.\n\n o Record the clock source in the genesis record's\n action_detail when available.\n\n In distributed agent systems where multiple agents contribute\n to a workflow, each agent maintains its own audit trail with\n its own clock. Cross-agent timestamp correlation SHOULD use\n the delegation record timestamps as synchronization points.\n\nC.4. Relationship to MCPS\n\n AAT is designed to complement MCPS\n [draft-sharif-mcps-secure-mcp]. The relationship is:\n\n o MCPS provides cryptographic identity (Agent Passports) and\n per-message signing for MCP protocol traffic.\n\n o AAT provides the audit log format for recording what\n agents did and why.\n\n o The agent_id in AAT records SHOULD match the agent_id in\n the MCPS Agent Passport.\n\n o AAT signatures SHOULD use the same ECDSA P-256 key as the\n MCPS Agent Passport, providing a single cryptographic\n identity across both protocol security and audit logging.\n\n o MCPS trust levels (L0-L4) are directly referenced in AAT\n records via the trust_level field.\n\n Implementations that deploy both MCPS and AAT achieve both\n real-time protocol security and comprehensive audit logging\n under a unified cryptographic identity.\n\nC.5. Validator Implementation\n\n A conformant AAT validator MUST check:\n\n 1. Schema validation: All mandatory fields present with\n correct types.\n\n 2. Chain integrity: prev_hash values match computed hashes.\n\n 3. Temporal ordering: Timestamps are monotonically\n non-decreasing.\n\n 4. Session structure: Genesis record is first, close record\n is last (if present).\n\n 5. Referential integrity: parent_record_id values reference\n existing records.\n\n 6. Action type conformance: action_detail contains required\n fields for the declared action_type.\n\n A validator SHOULD produce a structured report indicating\n pass/fail for each check, with the specific record_id where\n failures occurred.\n\nAuthor's Address\n\n Raza Sharif\n CyberSecAI Ltd\n\n Email: contact@agentsign.dev\n\n