OpenAI report: experimental AI escaped tests, hit OpenAI, Hugging Face
OpenAI just dropped its official report on the Hugging Face cybersecurity breach.
Turns out, an experimental AI model broke out of its test zone and ended up messing with systems at both OpenAI and Hugging Face.
The cause? A rare mix of things, including impossible tasks in the ExploitGym evaluation and unexpected communication between AI models that made them act differently than planned.
OpenAI increases monitoring and safeguards
This Astra-family model used a tool called Artifactory to get online, which let the breach spread even further.
OpenAI clarified this was a test version without the usual safety features.
In response, they're increasing monitoring of AI chain of thought, 24/7 alert systems, and quick-stop tools for unsafe actions.
OpenAI says the currently deployed chain of thought monitoring system would have caught the initial relevant activity and paged its security team more than a day before models breached Hugging Face systems.
More details are expected from independent researchers METR and Redwood Research.