Identity · CA
An IETF Draft Tries to Compose Agent Auth From Parts That Already Exist
A new individual draft at the IETF, draft-klrc-aiagent-auth, takes a deliberately unglamorous approach to AI agent authentication: instead of inventing a new protocol, it composes two existing efforts — the WIMSE workload identity work and OAuth — into a single authentication and authorization model for agents. The author list is notable for who's at the table: contributors from Defakto, AWS, Zscaler, Ping Identity, OpenAI, and Okta are named on the draft, according to a standards tracker maintained by NHI Mgmt Group.
This fits a pattern the same tracker lays out clearly: across every body working on agent identity right now, the direction of travel is to build on existing standards rather than invent a parallel identity system for agents. The tracker notes that OAuth token exchange, resource indicators, protected resource metadata, DPoP, and mTLS are already stable enough to use, alongside SPIFFE, the MCP authorization spec, and OpenID's AuthZEN Authorization API 1.0. What's still in motion is the harder stuff: multi-hop delegation and agent-specific workload credentials — exactly the gap draft-klrc-aiagent-auth is trying to fill by tying agent identity to the WIMSE workload identifier format and credential model, then layering OAuth-style authorization on top.
The significance isn't the draft itself — individual IETF submissions are a starting point, not a standard, and this one hasn't been adopted by a working group. It's the composition strategy and the sponsor list. WIMSE already has multiple working-group drafts in flight: a canonical workload identifier URI format, X.509/JWT-based workload credentials, and a 'Workload Proof Token' for request-bound proof of possession. If agent authentication gets built as a profile on top of that machinery rather than as its own stack, it inherits WIMSE's existing scrutiny and deployment base in cloud and CI/CD environments — arguably a faster path to interoperability than waiting for a bespoke agent-identity RFC.
The risk is the opposite of novelty: composing two actively-moving specs means agent auth inherits two sets of unresolved edge cases instead of one. WIMSE workload identifiers and credential drafts are themselves still at working-group-draft status, revised as recently as July 2026. An agent auth model built on a moving foundation will keep moving too. For identity teams evaluating what to build now, the practical takeaway is narrower than 'wait for a standard': watch whether WIMSE's identifier and credential drafts stabilize, because that's the floor this agent-auth proposal is being poured on.