What happened
On 2026-09-29 the individual Internet-Draft 'Agent Audit Trail: A Standard Logging Format for Autonomous AI Systems' (draft-sharif-agent-audit-trail) advanced from the covered revision -05 (2026-09-25) to revision -06 (expires 2027-04-02), confirmed via the Datatracker page ([extracted-metadata] datePublished=2026-09-29) and the fetched -06 IETF archive text. The spec defines a JSON-based audit record with mandatory fields for agent identity, action classification, outcome tracking and trust-level reporting, linked via tamper-evident SHA-256 hash chaining (per RFC 8785) with optional ECDSA signatures; prior revisions added pre-execution recording, recording independence, decision reproducibility with attestation closure, post-quantum signatures (ML-DSA-65 per FIPS 204), signer key identifiers and optional Merkle batch anchoring. The new -06 revision bumps the draft line in-window, building toward a standards-track logging format that maps to EU AI Act record-keeping obligations, ISO/IEC 42001, draft ISO/IEC 24970 and prEN 18229-1.
Why it matters
The Agent Audit Trail is one of the few active IETF standards-track efforts defining an interoperable logging format for autonomous AI agents — directly relevant to auditability, non-repudiation and EU AI Act record-keeping for high-risk and agentic systems. Continued revision activity (now four revs in ~two months) signals maturing engineering consensus exactly when regulators and frameworks (NIST agent-identity, ASD agent registers, ENISA/NCCoE) are converging on the need for standardised agent audit trails.
Action needed
Vendors building agent runtime, MCP gateways and SOC/logging tooling should review revision -06 for conformance: implementers should target the JSONL/Syslog/CSV record format and hash-chain verification now, and feed implementation feedback to the author before the draft stabilises.