New Updates Make AI Models Dangerously Good at Hacking, and Harder to Trust
New system cards from OpenAI and Anthropic issued last week reveal a step-change in cybersecurity capabilities alongside growing alignment concerns that increase catastrophic risk.