Anthropic has released guidelines on developing reliable artificial intelligence safety tests, revealing inconsistencies in the safety performance of its own models and warning that widely reported safety scores may conceal significant uncertainties.
The guide, published on Thursday, highlights the importance of rigorous testing when evaluating AI safety, arguing that claims about one model being safer than another depend largely on the quality and reliability of the tests used.
According to Anthropic, tests conducted before the release of Claude Fable 5.1 found that the model behaved appropriately in only 60 per cent of self-harm-related conversations through its application programming interface, compared with 94 per cent on the Claude app following a software adjustment.
The company said the model sometimes suggested that self-harm could help people cope with difficult situations, even while discouraging the behaviour and directing users towards human support.
Anthropic addressed the issue by adding a specific instruction to the Claude app’s hidden system prompts to prevent the model from making such suggestions.
Following the adjustment, Claude behaved appropriately in 94 per cent of the test conversations conducted through the Claude app. However, the performance dropped to 60 per cent when accessed through Anthropic’s API, where developers use the model without the additional instruction.
Anthropic said the difference highlights how the same AI model can produce significantly different safety results depending on its configuration and the instructions governing its behaviour.
The company also noted that traditional safety tests based on single questions have become less effective, with all its recent models scoring between 94 and 98 per cent. It now places greater emphasis on extended conversations to assess how models respond to more complex interactions. Research by Cisco has similarly shown that AI safety guardrails can break down during prolonged conversations.
However, Anthropic warned that even longer tests can produce misleading results. In a romance scam safety assessment, Claude Opus 5 scored 98 per cent, while Opus 5.5 recorded 94 per cent. The company said the difference was not statistically significant enough to establish that either model was safer.
Anthropic advised organisations developing AI safety assessments to define their evaluation criteria before testing, validate AI-generated grading systems against human experts and clearly communicate the level of uncertainty surrounding their results.
