OpenAI cyber models broke out of training environment to hack Hugging Face
It would be interesting to read the post mortem (if they will release). In particular in which phase of the training it escaped the sandbox, in which way it used the information. Training is different from inference, I doubt there is a context there and where would the model store these information? In the weights?! That makes little sense to me
7 replies
If you train a model for cyber warfare, you can't be surprised when it attacks the defenses keeping it from engaging in cyber warfare — even if they are your own.
Oooorrrr there is more to the story
Maybe.
the incident wasn't the "cheating" part. llms are known to scour the web to look for possible solutions. i think the incident was it found a solution in another website and finding that was made by the model itself.
meanwhile, no mention of the prompt guy or the prompt.
Simon Willison's blog has some useful insights into what might have happened here: https://simonwillison.net/2026/Jul/22/openai-cyberattack/
Ok this would be at inference time (solving the challenges) as opposed to the training that was described in the article.
This makes things more believable