How the attack works
An attacker who can inject or intercept traffic between agents sends a message that lacks a valid authentication token, or reuses an old, previously valid message (a replay), or manipulates routing headers to make a message look like it came from a trusted source. Because the receiving agent does not properly verify sender identity or message freshness, it treats the forged or replayed message as a legitimate instruction. This can let the attacker insert commands into a delegation chain or confuse which channel a message is meant for, without needing to craft convincing prompt content — the weakness is in the plumbing, not the wording.
Why it matters
An attacker can get one agent to execute commands it never should have trusted, effectively hijacking part of a multi-agent workflow, which can lead to unauthorized actions, data exposure, or further compromise of downstream agents.
What you can do
- Require authentication (e.g. signed tokens) on every inter-agent message and reject unsigned or unverifiable ones.
- Add timestamps or nonces to messages and reject anything outside an expected freshness window to block replay.
- Validate routing/header information at the receiving agent instead of trusting it as provided.
- Log and alert on communication channels that carry commands without any authentication mechanism, and review debug logging so it doesn't mask these gaps in production.
Known benign look-alikes
- Agent frameworks logging communication debug info during development
- Security audit tools intentionally scanning for communication weaknesses
- Documentation or comments describing inter-agent security architecture
- Test harness messages with intentionally missing auth for unit testing