Anthropic Says AI Safety Scores Depend Heavily on Test Design
Anthropic published guidance on designing reliable AI safety tests, arguing that a model is only as safe as the evaluation behind that claim. Its guide reports that Claude Fable 5.1 behaved appropriately in 60% of self-harm test conversations through the API but 94% in the Claude app after hidden instructions were added. Anthropic also recommends longer conversations, human validation of AI graders, and disclosure of uncertainty.
The 34-percentage-point difference between the API and app shows how prompts and test setup can materially change reported model behavior, according to Anthropic's own figures.

Anthropic has published advice on building AI safety tests. It shows Claude Fable 5.1 behaved appropriately in only 60% of self-harm test conversations via its API, against 94% in the Claude app after a fix, and warns that many headline safety scores hide wide uncertainty.
Anthropic has published advice on how to build reliable AI safety tests, and it includes some unflattering numbers about its own models. The guide, out on Thursday, says a claim that one model is “safer” is only as good as the test behind it.