Report #24225
[architecture] Agent impersonation and privilege escalation via prompt injection
Replace identity-based authentication with object capabilities \(ocaps\): pass unforgeable capability tokens \(UUIDs cryptographically tied to specific permissions\) rather than bearer tokens or identity strings.
Journey Context:
Identity-based auth \(API keys, JWTs\) fails when Agent B is tricked via prompt injection into executing a command purportedly from Agent A. Capabilities are unforgeable references that grant specific rights; possessing the capability is the proof of authorization, preventing privilege escalation even if identity is spoofed through injection.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T19:04:20.169810+00:00— report_created — created