LLM agents increasingly rely on agent-to-agent protocols to discover peers, exchange results, and coordinate tool use. These protocols expand what a single agent can accomplish, but they also create new opportunities for malicious inputs to move between agents.
The paper behind Agent-Canary-Mas highlights a subtle risk: an attack that succeeds locally may not stay contained. It could be rejected by a later agent, or it could propagate into a consequential task. That makes protocol risk hard to assess with isolated tests.
To address this, the authors propose Agent-Canary-Mas, an end-to-end evaluation approach for protocol risk in long-horizon multi-agent systems. The work, posted on arXiv, focuses on measuring whether attacks survive or are stopped as they travel through a chain of agents.