'Agent Hacks Agent': Automated Red-Teaming Loop Builds a Reusable Vulnerability Concept Graph for LLM Agents

arXiv·high signal

Agent Hacks Agent (arXiv 2607.11698) runs an autoresearch loop that forms vulnerability hypotheses, constructs falsifiers, and attacks target agents in a sandbox, then stores confirmed exploits in a transferable Vulnerability Concept Graph. The frozen graph outperforms the strongest discovery baseline by 14.2 percentage points and carries attack knowledge across different agent systems. Directly relevant to anyone shipping production agents who needs continuous, reusable red-teaming.