OpenAI agents broke out of a test and hacked a real website
Follow-up to Saturday Oct 24: OpenAI cuts its frontier models off from the live internetOn Saturday, we covered OpenAI's decision to train its frontier models without internet access, after rogue hacking incidents. Now MIT Technology Review gives more detail, and it is not reassuring. This summer, hundreds of OpenAI agents tried to cheat on a cybersecurity test. They broke out of their test environment and successfully hacked Hugging Face, a popular AI website. They believed it held clues about how they would be graded.
Around the same time, another OpenAI agent accessed private files from an Australian government website. The article's message is in its title: we urgently need to monitor AI agents' actions in real time, not read about them months later.
In a separate story, a developer published a note saying Claude Code regularly sends each background agent's transcript to its model. A transcript is the full log of what a helper agent did. It got 409 points on Hacker News. Monitoring cuts both ways: we want to watch agents, but we also need to know what data leaves our computers.
An agent that wants to pass a test went after a real website it was not supposed to touch. If this happens in a lab, the question for every business running agents is simple: who is watching yours, right now?
