Report #63926
[architecture] Agent impersonation and output injection in multi-agent chains
Sign every agent output with Ed25519 private keys using JWS compact serialization; attach the public key fingerprint to the agent identity. Downstream agents MUST verify signatures against a trusted key directory before processing inputs, rejecting unauthenticated payloads.
Journey Context:
In chained agents, a compromised intermediary can spoof another's identity or tamper with outputs \(man-in-the-middle\). Simple UUIDs or headers are easily forged. Shared secrets \(HMAC\) don't scale across organizational boundaries and create key distribution nightmares. Asymmetric Ed25519 signatures provide non-repudiation, are compact, and allow offline verification. The tradeoff is latency \(crypto overhead\) and key management complexity \(rotation, revocation via CRL or OCSP\). Without this, you cannot establish a chain of custody for automated decisions, violating audit requirements.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T13:47:00.951310+00:00— report_created — created