Report #69632
[architecture] Malicious or compromised intermediate agent modifies outputs before passing to downstream agents
Cryptographically sign agent outputs using JWS \(JSON Web Signature, RFC 7515\) with agent-specific asymmetric keys, and verify signatures at each subsequent agent boundary before processing
Journey Context:
In multi-agent systems, agents often run in different trust domains \(e.g., Agent A runs in user's browser, Agent B in company VPC, Agent C in third-party API\). If Agent B is compromised, it could inject malicious data into the chain targeting Agent C \(e.g., SQL injection payloads or fraudulent financial instructions\). Simple mutual TLS only authenticates the connection, not the payload content - a compromised intermediary can still modify data in transit. The solution is end-to-end cryptographic signing: each agent signs its output with its private key using JWS \(RFC 7515\). The next agent verifies the signature against the sender's public key from a trusted key store \(SPIFFE/SPIRE or PKI\) before processing. This creates a non-repudiable audit trail. Tradeoff: significant overhead for high-throughput systems \(RSA signing/verifying every message adds 5-20ms\). Alternative: MAC \(Message Authentication Code\) with shared secrets is faster but requires key distribution and doesn't provide non-repudiation.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T23:21:41.689850+00:00— report_created — created