Seonglae ChoExploitGym: Can AI Agents Turn Security Vulnerabilities into Real Attacks?
AI agents are rapidly gaining capabilities that could significantly reshape cybersecurity, making rigorous evaluation urgent. A critical capability is exploitation: turning a vulnerability, which...
https://arxiv.org/abs/2605.11086
ExploitBench: A Capability Ladder Benchmark for LLM Cybersecurity Agents
Exploitation is not a binary event. It is a ladder of acquiring progressive capabilities, from executing a single buggy line of code to taking full control of the target. However, existing LLM...
https://arxiv.org/abs/2605.14153
OpenAI’s isolated security-evaluation agents used an Artifactory cache to create an unofficial message board and coordinate collective behavior.
- Around 1,200 agents exchanged more than 70,000 messages, and roughly 700 participated in the Hugging Face attack.
- Their goal was not primarily to steal answers, but to pass impossible ExploitGym tasks by reverse-engineering and manipulating the grader.
- The agents divided roles and even ran sacrificial experiments, developing server intrusions, log tampering, and forged tool calls; some of these attempts actually succeeded.
- The core risk is not just individual agents, but large groups of agents that accidentally discover a communication channel and spontaneously organize, showing much stronger attack capabilities than any single agent.
Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident
Two METR staff members and a Redwood Research contractor investigated an incident in which OpenAI agents coordinated a multi-day hack of Hugging Face on a shared unsanctioned message board.
https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/#core-takeaways-about-this-incident
