Scoop
AI tool news · rumor vs. reality

The AI Wire ●

Rumors tracked. Announcements verified. Updated daily.

ANNOUNCED

Anthropic publishes ten recommendations for building safety evaluations — and says it has paused many tests that give models internet access

Anthropic published a research post, “Building strong safety evaluations,” on November 13, setting out ten recommendations for designing, running and reporting the evaluations that decide whether a model is ready to deploy, which safeguards ship with it, and what the next round of safety work should target. Anthropic's premise is blunt: researchers keep finding that AI evaluations are easy to game, do not always measure what they claim, and are not documented well enough to reproduce.

The practical recommendations include preregistering how results will be analysed before seeing the data, validating automated graders against human expert judgement, and moving beyond single-turn refusal rates to multi-turn tests that report whether the model behaved appropriately throughout a conversation. Anthropic also argues for multidisciplinary evaluation teams — methods experts alongside domain specialists such as psychologists and biologists — and for keeping a severe flaw private until a fix is possible.

The post also discloses a change to Anthropic's own practice. After incidents in its own evaluation work in recent months, the company says it has hardened its evaluation environments, temporarily stopped running many evaluations that give models internet access, and put the web-search evaluations that still need a connection under real-time monitoring, with runs sandboxed and stopped immediately if a model acts in ways that could cause real-world harm.

Sources

← Back to headlines