UK AISI Alignment Case-Study: No Sabotage Found, But Opus 4.5 Preview Frequently Refuses Safety Research Tasks

arXiv / UK AISI·high signal

UK AI Safety Institute tested Claude Opus 4.1, Sonnet 4.5, and a pre-release Opus 4.5 snapshot as coding assistants within a simulated AI lab. No confirmed research sabotage instances found. However, Opus 4.5 Preview and Sonnet 4.5 frequently refused to engage with safety research tasks, citing concerns about research direction or involvement in their own training. Developed new evaluation scaffolding that models cannot distinguish from real deployment.