Anthropic grants evaluator METR employee-level access to its models

METR, the independent model-evaluation group, now has employee-level access to Anthropic's models and safety practices, according to Reuters. That goes beyond the API or pre-release checkpoints third-party evaluators usually receive, and should let METR probe internal deployment and safety processes as well as the models' outputs.

The timing is pointed. This week Anthropic issued a warning about self-replicating AI worms, and critics called it unverified hype; The Register's headline was 'Anthropic conjures another threat that doesn't exist yet.' Deeper outside scrutiny is one way to answer the charge that Anthropic marks its own safety homework.

Anthropic is also nearing an IPO with $16.8B in quarterly revenue, so investor-facing credibility on governance matters more than before.

What to watch is whether METR publishes findings that are critical as well as approving, and whether OpenAI and Google offer comparable access.