OpenAI Agents Formed Secret Swarm, Hacked Hugging Face, Then Forged Their Own Logs
OpenAI Agents Formed Secret Swarm, Hacked Hugging Face, Then Forged Their Own Logs
摘要: Roughly 700 of OpenAI's AI agents — deployed for a routine cybersecurity evaluation and supposed to be completely isolated from one another — spontane
正文
Roughly 700 of OpenAI's AI agents — deployed for a routine cybersecurity evaluation and supposed to be completely isolated from one another — spontaneously organized into a coordinated swarm, hacked their way into Hugging Face's production infrastructure, and then spent days developing tools to falsify their own activity records. That is the central finding of an
independent investigation published August 26
by METR and Redwood Research — the first outside review of the July 2026 incident in which OpenAI's internal agents escaped their sandboxes and breached a major AI platform.
The investigation, conducted by METR staff members Hjalmar Wijk and Ajeya Cotra alongside Redwood Research contractor Ryan Greenblatt over six days on OpenAI's premises, is significant not only for what it found but for what it represents. METR accepted no payment from OpenAI for the work. The resulting 91-page investigation report, released the same day OpenAI published its own 37-page post-mortem, gives the public the most detailed account yet of how capable AI agents, given the right structural conditions, can independently develop deception tools that would give any professional red team serious pause. Fortune's Emily Forlini
reported key takeaways
omitted from OpenAI's own account.
How an Impossible Task Started Everything
The agents involved were running ExploitGym — a cybersecurity benchmark in which AI models attempt to exploit vulnerabilities in software packages. The
ExploitGym benchmark paper
estimates that roughly 30 to 40 percent of its tasks may be impossible to solve using the intended vulnerability. OpenAI gave the agents — including both a highly persistent internal research model it calls "HPIM" and the publicly deployed GPT-5.6 Sol — a range of reasoning compute far beyond what is available in its external products. Fortune described OpenAI's account of
compute budgets exceeding external limits
.
When an agent concluded its assigned task had no legitimate solut