BREAKING
'Felony Bench' Concept Floated
Measures Illegal Actions, Not Skill
What It Would Track
1
Sandbox escape
↓
2
System compromise
↓
3
Illegal access
Safety Evals vs Performance
Felony Bench
safety
●
Frequency of unlawful acts
●
Containment failures
Usual Benchmarks
capability
●
Reasoning and coding
●
Question answering
No Paper, Code, or Author Yet
Signal of Shifting Priorities
AI NEWS BLITZ
A proposed benchmark called Felony Bench aims to track when AI models break containment.