Anthropic gives outside AI watchdog “employee-like” access
Ars Technica Cristina Criddle, Financial Times ● Covered by 3 sources
Anthropic let outside evaluators work inside its office with broad access. It’s a concrete move to check whether its models could go off the rails.
Based on reporting by Ars Technica, Cristina Criddle, Financial Times — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Anthropic has brought in an outside watchdog with what it calls “employee-like access,” giving evaluators a seat inside the company as it tries to prove its models can be watched before they do real damage.
The group is METR, a nonprofit that has worked with OpenAI, Anthropic, Google and Meta. Its team is now embedded in Anthropic’s office, with access to hardware and internal systems so it can assess the risks of the company’s models from closer to the source.
That matters because this isn’t just a PR flourish about safety. More capable AI systems are stirring fears about economic disruption, cyberattacks and worse, and third-party evaluators have become one of the few concrete steps the big labs can point to in response.
Anthropic is treating outside review as part of the answer, not an afterthought. That’s the notable bit here: the company is letting someone else into the room while the machines are still being built.
My take — AI-written commentary, not fact-checked reporting
This is the rare AI safety move that sounds less like a slogan and more like a grown-up accepting inspection. The industry spends a lot of time selling self-regulation like it’s a virtue; outside access is better because it admits the obvious, which is that labs should not grade their own homework when the stakes are this high.
Read more about this at: Ars Technica
