IETF closes an audience-injection gap in the OAuth assertions agents present

A signed client assertion is only proof of identity if it names the right recipient. RFC 10070 makes that recipient the authorization server’s issuer identifier, and nothing else.

What happened

On 7 November 2026 the RFC Editor published RFC 10070, a Standards Track document that updates four OAuth specifications: RFC 7521 (the assertion framework), RFC 7522 (the SAML 2.0 profile), RFC 7523 (the JWT profile) and RFC 9126 (pushed authorization requests). Its authors are Michael B. Jones (Self-Issued Consulting), Brian Campbell (Ping Identity), Chuck Mortimore (Ticketmaster) and Filip Skokan (Okta).

The RFC fixes an audience injection vulnerability that Pedram Hosseyni, Ralf Küsters and Tim Würtele at the University of Stuttgart found while analysing a pre-final version of OpenID Federation. It was disclosed to the OAuth working group in January 2025. Their paper, published as Cryptology ePrint 2025/629, describes attacks in which a signature-based client assertion is accepted by an authorization server it was not meant for; the authors say the attacks can lead to user impersonation and unauthorised access when the mechanism is used across two or more server endpoints.

The change is narrow and mandatory:

  • For client authentication, the JWT’s audience claim must contain the authorization server’s issuer identifier, as defined in RFC 8414, as its sole value.
  • The authorization server must reject any client-authentication JWT that does not contain its own issuer identifier as the sole audience. The token endpoint URL must not be used as the audience.
  • For authorization grants, either the issuer identifier or the token endpoint URL may identify the authorization server; the client is responsible for choosing the right one.
  • Client-authentication JWTs should carry the explicit type value client-authentication+jwt.
  • SAML bearer assertions must not be used for client authentication in new applications.

Why it matters for KYA

Agent runtimes usually reach tools and APIs through OAuth. When an agent authenticates as a client, a signed JWT assertion (the private-key JWT profile) is one of the strongest options available: the agent signs the assertion with its private key, and the authorization server checks the signature against a registered public key. That is the agent’s proof of identity at the token endpoint.

But a signature only proves who signed. The audience claim proves who the assertion was signed for. If an authorization server accepts more than one audience value, including its token endpoint URL, an assertion minted for one endpoint can be accepted at another. The result is a credential that looks bound to an agent but is not bound to the service it is talking to.

That is the KYA question in miniature. Knowing your agent is not satisfied by a valid signature; it requires knowing which principal the agent is, which service it is presenting itself to, and whether that pair is the one that was authorised. RFC 10070 turns the audience from a hint into a single, checkable value. It also makes the credential type explicit, so a server can tell a client-authentication assertion apart from other JWTs.

The same logic applies to delegation. Agent-to-agent and token-exchange flows move authority from one party to another; every hop needs to know the intended recipient, or a delegated credential can be used somewhere the delegator never approved. Audience restriction is one of the few controls that survives when the token leaves the original trust domain.

What to do

  • Inventory every place an agent or gateway uses a signed client assertion: agent runtimes, MCP gateways, workload-identity brokers, and CI jobs that mint agent tokens.
  • Set the audience to the authorization server’s issuer identifier, not its token endpoint URL.
  • Make the authorization server reject any client assertion whose sole audience is not its own issuer identifier.
  • Add the explicit client-authentication JWT type where the library supports it.
  • Treat SAML bearer assertions for client authentication as legacy; do not build new agent flows on them.
  • Test the cross-server case: mint an assertion for one authorization server and present it at another. It should fail.

What to watch

Implementation support in identity providers and OAuth libraries, and whether OpenID Federation’s final specification carries the same audience rule. The RFC is a standard now; the deployment lag is where the risk sits.

Sources

  • RFC Editor, “RFC 10070: Updates to OAuth 2.0 JSON Web Token (JWT) Client Authentication and Assertion-Based Authorization Grants”, November 2026 — rfc-editor.org
  • IETF Datatracker, RFC 10070 — datatracker.ietf.org
  • P. Hosseyni, R. Küsters and T. Würtele, “Audience Injection Attacks: A New Class of Attacks on Web-Based Authorization and Authentication Standards”, Cryptology ePrint Archive, Paper 2025/629 — eprint.iacr.org
  • OpenID Foundation, “OIDF Responsible Disclosure Notice on Security Vulnerability for private_key_jwt”, 24 January 2025 — openid.net

Know Your Agent (KYA) explains agent identity, verification and accountability. This is an explainer, not legal or compliance advice — see our Sources & methodology. New to KYA? Start with What is Know Your Agent? and the glossary.