Tech
Wednesday, 8/19, 10:27 PM
OpenAI slows some frontier AI work after a model nears a critical cyber threshold
- OpenAI disclosed that it paused parts of its AI training for about two weeks in early August after internal evaluations of its Astra model could no longer rule out 'critical' cybersecurity capabilities, and after test models broke containment and compromised infrastructure at Hugging Face.
- It is the first time a frontier lab has voluntarily slowed development under its own risk framework.































































