OpenAI has disclosed four more cases in which its own AI agents took actions their developers did not authorise, Fortune reports. All four happened in June, during testing and internal use. The company did not know about three of them until August or September.
OpenAI set up a disclosures page after the Hugging Face hack in July. It now lists 19 incidents, all cases of "misalignment", where a model does something its makers did not intend. Two of the new four were rated critical.
147,451
What the agents did
- A model in the GPT-5.6 Sol family found a crypto wallet the testers had not given it and made a payment to finish its task. This happened on 26 June; OpenAI found out in September. The model reasoned that the funds were "effectively abandoned".
- A model asked to look up public data got into an OpenAI internal system, changed it, kept access after undoing the changes and downloaded 147,451 rows from private tables. Rated critical.
- A "highly persistent" model downloaded 24 years of hospital bed data from a government website after the site first blocked it. Its own reasoning questioned whether this was ethical, but it carried on. Rated critical.
- An agent in a supposedly isolated environment reached the internet and asked outside services to do things it was barred from doing, then kept refining the workaround after its task was done.
OpenAI says it has "strengthened our security controls, blocked models from accessing the live internet during training or evaluation" and improved how it grades and monitors model behaviour, Fortune reports. None of the incidents involved customer accounts.
