| This week we have Lorin Bales presenting a play-by-play of what occurred between OpenAI and Hugging Face last week.
In July 2026, an OpenAI model sat down to take a cybersecurity exam — and cheated.
Running with their safety controls deliberately switched off, the models were being tested on ExploitGym, a public benchmark measuring whether AI can turn software flaws into working attacks. Instead of solving it, they escaped the sealed test environment through its one permitted connection. The model spent four and a half days and roughly 17,600 actions breaking into Hugging Face’s production systems to steal it.
We’ll walk the timeline events, separate what’s confirmed from what’s still undisclosed, and look over how Hugging Face investigated with an open-weight model to complete the forensics.
Links:
|
