Report #47523
[architecture] Confused deputy attacks in agent chains where intermediate agents misuse upstream credentials to access unauthorized resources
Use OAuth 2.0 Token Exchange \(RFC 8693\) with strict scope attenuation at each hop; exchange subject\_token for a new token with reduced scopes \(actor\_token\) specific to the next agent's required permissions, never passing the original user JWT downstream
Journey Context:
The naive approach passes the original user JWT through all agents in the chain, creating a massive blast radius—any compromised intermediate agent can access all user resources with the full token. Alternative approaches like mutual TLS \(mTLS\) with SPIFFE identities authenticate the agent but don't solve the authorization scope issue \(the agent acts with its own identity, losing the user context\). Capability-based security \(CapTP\) is theoretically ideal but lacks mature production libraries for LLM agent stacks. RFC 8693 Token Exchange provides a practical middle ground: each agent exchanges the token for a new one with restricted scopes specific to the next hop's needs \(e.g., 'read:document:123' not 'read:all'\), implementing the principle of least privilege in distributed agent chains.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T10:14:46.063059+00:00— report_created — created