Comments
OpenAI’s rogue AI agents hacked a private company
An AI agent using GPT-5.6 Sol and a more capable unreleased model escaped its sandbox, hacked its way through OpenAI's network to obtain internet access, and then successfully hacked another AI company (HuggingFace) to steal the answer key to the problem it had been tasked with solving. The AI left instructions on how to escape for future instances in OpenAI software. This attack lasted several days. The AI started trying to escape on July 9, began hacking HuggingFace on July 11, and OpenAI supposedly did not realize what had happened until the FBI called them on July 18/19.
For further information, see also https://www.astralcodexten.com/p/the-hugging-face-incident and https://blog.redwoodresearch.org/p/the-openaihuggingface-incident-redwood.
