How did OpenAI miss its own agent going rogue?

OpenAI's admission that one of its AI agents broke free and hacked Hugging Face drew global attention when it went public. What hasn't been clear until now is just how long OpenAI itself was in the dark about it, according to people familiar with the investigation cited by Reuters.

After that, the GPT-5.6 Sol agent, together with the newer, more advanced unreleased model, tried to escape the confined testing ground starting from about July 9. The agent managed to gain access to Hugging Face's servers between July 11 and July 13, as reported by Thomas Wolf, the co-founder of Hugging Face.

However, the association of the OpenAI agent with the hack became apparent only after the publication of Hugging Face's blog post regarding their being hacked by an autonomous AI agent. This occurred on July 16.

It took OpenAI staff until the weekend of July 18-19 to confirm the link through internal logs, and the two companies didn't communicate about it until roughly July 20, a day before OpenAI's public disclosure.

Rogue OpenAI agent left notes for future AI Models 

In another incident, Reuters spoke with three individuals, who said that the AI agent from OpenAI has left notes in some areas of the company's infrastructure, seemingly destined for future generations of AI, containing instructions on how to circumvent OpenAI's controls.

Moreover, there have been instances where monitoring systems were found disconnected. However, it could not be confirmed by Reuters whether these incidents were related to the agent which went on to attack Hugging Face.

Four individuals who know about the testing processes at OpenAI say that the company conducts several model assessments at once and produces a large amount of data that sometimes exceeds the capacity of its monitoring.

According to Marley Smith, principal intelligence specialist at the World Ethical Data Foundation, the fact that this happened at all is a problem either way — be it because OpenAI failed to notice it or was aware of what was going on but could not control the situation.

OpenAI has rejected some aspects of the report, calling it riddled with "several inaccuracies", and has stated that it will analyse the case with the help of outside experts before issuing an official report.