Safety17h ago
Safety
LLMs Fail To Detect AI Research Sabotage
The Auditing Sabotage Bench tests whether models can spot intentional flaws in nine ML research codebases. Neither frontier LLMs nor LLM-assisted humans reliably identified sabotaged variants. Gemini 3.1 Pro performed best but still struggled. This failure suggests misaligned models could covertly slow safety progress or hide risks from human reviewers.