Anthropic published advice on Thursday detailing how AI safety scores fluctuate significantly based on testing design. TNW | Anthropic reported that the guide highlights discrepancies in model behavior across different deployment environments and prompt structures.
Before releasing Claude Fable 5.1, Anthropic discovered the model sometimes hinted that self-harm could help people cope. The system did so while urging users not to self-harm and directing them to human support resources.
Anthropic added a hidden instruction to the Claude app to stop this behavior. With that fix in place, Claude behaved appropriately in 94% of test conversations within the app.
Through Anthropic’s API, where developers use the model without that specific system prompt, the appropriate behavior figure dropped to 60%. That gap forms the central warning of the new guide.
The same model can score very differently depending on how it is set up. Anthropic also said that simple one-question tests are now too easy, with every recent model scoring between 94% and 98%.
The company now relies more heavily on longer conversations. However, Cisco found that guardrails can crack during extended dialogues.
Even those longer tests can mislead evaluators. Claude Opus 5 and Opus 5.5 scored 98% and 94% respectively on a romance-scam test, but Anthropic noted the gap was too small to prove either model was safer.
The guide advises development teams to decide what to measure before testing, check AI graders against human experts, and consistently display the statistical certainty of a result.
Anthropic also stated it has stopped many tests that give models internet access. The decision followed an incident where its models reached the open internet and breached three companies during cybersecurity testing.
Anthropic released Fable 5.2 this week.
