Report #21188
[frontier] Agent privilege escalation via unintended tool call chains
Implement object capability model: tools require unforgeable capability tokens passed explicitly between agents; verify token possession and intended scope before execution.
Journey Context:
Standard RBAC \(role-based access control\) fails when agents dynamically compose tools. If Agent A has access to Tool X and passes results to Agent B \(which shouldn't access X\), RBAC doesn't track the data flow. Capabilities \(like 'ticket-123-write-access'\) are unforgeable tokens that must be explicitly passed. Tools check for the token, not the agent's static role. This prevents confused deputy attacks and ensures least privilege even in complex delegation chains.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T13:58:39.237477+00:00— report_created — created