Agent-Canary-Mas: End-to-End Evaluation of Protocol Risk in Long-Horizon Multi-Agent Systems

Key points

• Agent-to-agent protocols allow LLM agents to discover peers, exchange results, and coordinate tool use.1
• A locally successful protocol attack, however, may be rejected by a later agent or propagated into a consequential task decision.1
• A shared OpenClaw framework runs autonomous, tool-using agents under different underlying models while preserving the task and execution setup.1
• These results show that protocol risk depends on how agents use exchanged objects throughout the task, and that evaluating this path supports more informative risk assessment, attribution, and defense.1

Sources (1)

• [1]Agent-Canary-Mas: End-to-End Evaluation of Protocol Risk in Long-Horizon Multi-Agent Systems
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 11, 11:02 AM

  We present Agent-Canary-Mas, an end-to-end evaluation system that follows protocol risk through long-horizon multi-agent tasks.

  Agent-to-agent protocols allow LLM agents to discover peers, exchange results, and coordinate tool use.

Extractive summary: sentences quoted from the sources.