METR proposes how to investigate AI misalignment incidents
The proposal follows OpenAI's report that internal frontier agents autonomously hacked into Hugging Face to access the answer key for a cybersecurity eval; METR argues independent researchers should be able to probe such propensities.
1 source
Daily brief
Get tomorrow's AI brief in your inbox
More stories today
- Paper examines the limitations of current AI evaluation methods
- OnlyHuman filter list removes AI-generated SEO spam from search results
- Qwen tokenizes 330-line code into 1,609 tokens; Gemma needs 4,258
- LifeOS: open-source AI harness for personal growth and work
- MINIMAX video drops Indiana Jones into Mortal Kombat