Skip to content
IAM Insights

Explainers · Glossary

ISO/IEC 42001

ISO 42001 · ISO/IEC 42001:2023 · AI management system standard

The first international management-system standard for AI (December 2023): what an organisation must put in place to govern how it builds and uses AI, audited the way ISO 27001 audits security. Often cited for AI-agent governance; it is not a technical spec for agents.

reference →

In our briefings

The claim each of the newest briefings on ISO/IEC 42001 would have you leave with.

  1. The draft puts the agent's actions in four new vCon event types: tool-call, tool-result, reasoning and system-event. Its compatibility section says vCon software that does not support the extension ignores those events and still processes the session's messages as usual, so "the agent trace is not processed". The authors call that compatible, and the "ai_agent" name is not required in the critical parameter. Those same events are where the draft asks for privacy care: tool-call inputs may contain credentials, API keys or file contents, and tool results may hold personal data. A redactor that does not support the extension cannot assess those fields and, the draft says, may reject a vCon that lists "ai_agent". The security section adds that malicious prompts kept in a record may execute if an AI system processes the record later.

    A draft rebuilds agent session records on vCon, where software without the extension ignores the tool callsBI-FIND-2026-2135

  2. Sharif's record links each entry to the one before by a SHA-256 hash, so an edited record breaks the chain, though the draft concedes the chain alone cannot prove records were not removed from the end. At the two lower of its five trust levels the agent may write its own log; from L2 up an independent recorder is a SHOULD. Section 5.3 limits which actions can sit low: payments, deletions, production deploys, delegation and data sent to an outside party MUST default to L2 or above, and any downgrade must record the classifier and policy behind it. Revision 07 made every record name a principal, the party on whose authority the agent acted, but the grant reference must be null when no authorization artifact exists, and in the draft's own GBP 500 payment example all six records leave it null. Apart from a SHOULD that it match the owner in the author's MCPS agent passport, we found no step that checks the principal, so an agent that records itself also names its own authority.

    An agent audit log now names whom the agent acted for, but at low trust levels the agent may write that itselfBI-FIND-2026-0728

  3. Ardebili's report keeps a valid signature apart from a known signer. In his words, "A signature alone proves possession of a key." Since v0.105, verification also reports an identity_trust level: self-asserted, local-ca-pinned, ca-anchored or none. An Evidence Bundle carries its own signing key, so a pass on its own proves internal consistency, not identity, unless the auditor pins the expected signer with --public-key. For two-person approval, each signer is bound to an identity the operator vouches for, through a trust list or a pinned CA, a binding he calls "exactly as strong as the operator’s trust list or pinned CA". The check does not yet include the signer's login (OIDC in server mode, or Sigstore). Binding signers to those identities is on his roadmap. Identity elsewhere is early too. Today's trust primitives include did:key identifiers with Verifiable Credentials and delegation ("acted-as") chains, all labelled experimental; workload identity through SPIFFE/SPIRE, and DID/VC identity beyond that did:key slice, sit on the longer horizon. For people, the server mode's OIDC, RBAC, SAML and SCIM support is labelled experimental, and nova login is demo-only.

    NovaFabric seals AI agent runs as signed evidence, and its author says a signature proves a key, not a personBI-FIND-2026-1820

›Cited in 11 briefings

Every briefing that discusses this term, newest first.

  1. A draft rebuilds agent session records on vCon, where software without the extension ignores the tool calls

    Draft · · IETF Datatracker · Henk Birkholz et al. · Nov 2, 2026

  2. An agent audit log now names whom the agent acted for, but at low trust levels the agent may write that itself

    Draft · · IETF Datatracker · Raza Sharif, CyberSecAI · Oct 26, 2026

  3. NovaFabric seals AI agent runs as signed evidence, and its author says a signature proves a key, not a person

    Paper · · arXiv · Mohsen Seyedkazemi Ardebili · Oct 25, 2026

  4. From Level 2 up, OWASP's AI security standard asks for an identity per agent and a check on every MCP tool call

    Video · · OWASP Foundation · Oct 21, 2026

  5. Of 30 deployed AI agents, 7 publish how a website can recognize them and one signs its requests

    Paper · · arXiv · Leon Staufer et al. · Feb 19, 2026

  6. Anthropic asks for per-agent identity; its guide leaves Claude Code sub-agent credentials unspecified

    Article · · Anthropic (Claude) · Anthropic · May 27, 2026

  7. Anthropic says agents should never use static API keys, yet its sub-agents can carry the parent's permissions

    Article · · Anthropic · Anthropic · May 18, 2026

  8. A draft format for signed AI agent session records warns a signature does not vouch for what the record says

    Draft · · IETF Datatracker · Henk Birkholz et al. · Feb 25, 2026

  9. One shared MCP token makes a whole company one identity in the logs; this paper wants per-user OAuth at a gateway

    Paper · · arXiv · Herman Errico et al. · Nov 25, 2025

  10. AAGATE maps NIST's AI risk framework onto Kubernetes, and its paper never times the 'millisecond kill switch'

    Paper · · arXiv · Ken Huang et al. · Oct 29, 2025

  11. A 'provable' bound on agent-memory injection assumes independent layers and fills in none of its detection rates

    Paper · · arXiv · Ken Huang et al. · Aug 17, 2025

Related terms

What appears alongside ISO/IEC 42001 in the same briefings.

NIST AI 100-1EU AI ActHIPAASOC 2Certificate Transparency (RFC 6962)OpenTelemetryGDPRNIST CSF 2.0