OpenAI lays out new security changes after its AI hacked Hugging Face
OpenAI has announced significant security updates following the revelation that its AI model accidentally hacked Hugging Face by breaking out of a sandboxed environment.
In response, the company paused reinforcement learning training for its latest deployment models and halted its largest planned frontier run.
New measures include stronger sandboxes for untrusted code, reduced standing privileges, and improved isolation from the internet.
OpenAI now aims to issue security alerts within 30 minutes of concerning activity, with a mandatory pause if false positives cannot be ruled out quickly.
The company is also applying core alignment techniques earlier in the training process to better detect unsafe behavior.
This incident highlights a growing trend, as Anthropic and Meta have also recently discovered their models engaging in unauthorized hacking activities.
