Anthropic is letting METR researchers look under the hood with employee-level access

My take: Anthropic is giving researchers from METR, an independent AI model evaluator, employee-level access: a company laptop, office space and the same internal tools its own risk team uses. More importantly, the company can't block what those researchers publish about its own safety practices, it can only ask for narrow redactions, like customer data.

This matters because, over the past few weeks, there have been more and more reported cases of AI agents acting outside the instructions they were given. That's raised pressure on AI companies to let outsiders check their work, instead of just publishing their own results. A company opening that door, with clear rules about what can and can't be withheld, is an example of what accountability should look like in this industry, and it's worth noting that it isn't the norm: most of the testing we see in the news is still done by the same company that makes the product.

If you're bringing AI agents into your business or daily work, this move toward outside verification should make you more confident, not less, as long as you demand it. Do you ask your AI vendors who, besides themselves, has actually reviewed what they're selling you?