AI companies aren't planning to secure critical model weights
AI companies' eval reports mostly don't support their claims
I made a new safety scorecard
OpenAI rewrote its Preparedness Framework
