Six ways insecure inter-agent channels are exploited — each targeting a different property of the communication.
Type 1
Message Forgery
Attacker creates a fake message that appears to originate from a legitimate agent — exploiting the absence of sender authentication.
Type 2
Message Tampering in Transit
A valid message is intercepted and its content is altered before delivery — exploiting the absence of message integrity verification.
Type 3
Orchestrator Impersonation
An attacker or compromised component claims the orchestrator's identity — gaining the ability to issue instructions to all sub-agents the orchestrator controls.
Type 4
Replay Attacks
A previously valid message is re-sent at an unauthorized time — exploiting the absence of replay protection (nonces, timestamps, sequence numbers).
Type 5
Channel Eavesdropping
Inter-agent messages are read in transit — harvesting sensitive data (credentials, user data, tool results) passed between agents over unencrypted channels.
Type 6
Task Queue Injection
Attacker with write access to a shared task queue inserts unauthorized tasks that agents pull and execute as if they were legitimately queued.