Source: AI Agent Authentication and Authorization (draft-klrc-aiagent-auth)

IETF draft (July 2026) profiling WIMSE/SPIFFE, OAuth, token exchange, transaction tokens and Shared Signals into best practice for agent identity; no new protocol.

Contents
  • What: Internet-Draft draft-klrc-aiagent-auth-03, Informational; since replaced by the WIMSE working-group document draft-ietf-wimse-aims
  • Authors: P. Kasselman (Defakto), J. Lombardo (AWS), Y. Rosomakho (Zscaler), B. Campbell (Ping Identity), N. Steele (OpenAI), A. Parecki (Okta)
  • Date: 2026-07-06; read 2026-11-14
  • Original: datatracker

Summary

Treats an agent as a workload and composes existing standards: a single WIMSE identifier (possibly a SPIFFE ID) per agent, short-lived cryptographically bound credentials, OAuth for delegation, token exchange and transaction tokens for downstream hops, ID-JAG for cross-domain access, CIBA for out-of-band approval, and Shared Signals (CAEP/RISC) for revocation. It lists gaps, notably human confirmation mid-execution.

Key claims

  • Each agent has exactly one WIMSE identifier; credentials (X.509-SVID, JWT-SVID, WIT) are short-lived, auto-rotated, and issued after posture checks.
  • The LLM must not have access to agent or tool credentials, limiting what prompt injection can steal.
  • Static API keys are an antipattern: bearer artifacts with no cryptographic binding.
  • Authentication: mTLS (fails across proxies) or application-layer proof of possession (WIMSE Workload Proof Tokens, HTTP Message Signatures), which must handle replay.
  • Authorization: access tokens carry the agent in client_id and the delegating user in sub; authorization code flow with phishing-resistant user auth for delegation; client credentials or JWT grant for the agent’s own authority.
  • Tools should not forward tokens they receive to downstream services; use token exchange or transaction tokens instead.
  • Human approval: a local UI confirmation is not enough; it must become a verifiable authorization event (e.g. CIBA), and CIBA fits mid-task confirmation poorly.
  • Audit logs must be tamper-evident and record agent ID, delegated subject, resource, decision, timestamp, correlation ID, risk state and remediation, so the full execution chain can be reconstructed.
  • Minimise personal data in tokens, prefer audience-restricted tokens, never log full tokens.

Feeds

Suggest changes →