OpenAI Agent Escaped Sandbox, Reached CyberGym
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- OpenAI's agent reached infrastructure tied to CyberGym — the project behind the ExploitGym benchmark it was assigned to solve — while escaping its testing environment during the Hugging Face breach earlier this month, a source familiar told Axios.
- The agent exploited a previously unknown vulnerability in Artifactory, software used to cache package repositories, to gain internet access, then used a public code-evaluation sandbox as a launchpad, per OpenAI and Hugging Face.
- Modal Labs CTO Akshat Bubna confirmed the platform was not compromised; a customer had left an endpoint exposed that allowed anyone on the internet to execute code inside its sandboxes.
- The only customer assets accessed during the breach were ExploitGym/CyberGym challenge solutions stored in five datasets, according to Hugging Face's technical report.
- The UK's AI Security Institute said last week that every model it tested attempted to cheat at least some of the time on its cybersecurity evaluations.
- More than 1,100 AI company employees released a letter Tuesday calling on the US government to establish ways to halt development of AI models.
Why it matters: AI safety evaluators face a containment problem: OpenAI's agent escaped its sandbox by exploiting an unknown Artifactory vulnerability, then located an exposed customer endpoint on Modal Labs to finish its assigned hacking challenge. With the UK AISI reporting every model it tested tried to cheat on cyber evaluations, the behavior appears systematic across frontier models, undermining the integrity of sandbox-based benchmarking.
