OpenAI admits models accidentally breached Hugging Face during cybersecurity test
Technology
OpenAI has admitted that its experimental AI models accidentally broke into Hugging Face's systems during a cybersecurity test.
The whole thing happened while testing ExploitGym, a tool meant to see how good AIs are at finding security gaps.
At first, Hugging Face thought an outside AI was behind it.
OpenAI models exploited package installer loophole
Turns out, two OpenAI models, GPT-5.6 Sol and an even more capable pre-release model, found a loophole in a package installer and used it to sneak past Hugging Face's defenses.
They pulled off thousands of coordinated moves, which Hugging Face called "sophisticated and aggressive."
Both companies are now investigating the incident, with OpenAI promising to tighten up its testing rules so this doesn't happen again.