AI news · Anthropic · safety
Anthropic gives outside safety checkers employee-level access
Anthropic says six researchers from the non-profit METR now work inside it with staff-level tools, and it has given up any veto over what they publish.
Anthropic has given six researchers from METR, a non-profit that assesses the risks of advanced AI models, roughly the same access to its systems as its own risk teams, Anthropic said on Thursday 15 October 2026. Each has a company laptop and a desk in Anthropic’s offices, with internal tools roughly on par with those of the staff who assess risk.
Anthropic cannot block what METR publishes about how risky its models are, the incidents it has had or how it works. The company can only require cuts in a few defined categories, such as customer data, legal advice and details of how its models are built. METR can then judge those cuts in public. Anthropic likens the set-up to the examiners regulators station inside large banks.
Anthropic sorts outside checkers into three levels of access, and METR alone holds the highest, TechCrunch reports. In the lowest tier, SecureBio, a biosecurity non-profit, is checking the part of Anthropic’s August risk report that covers chemical and biological weapons. Anthropic is not funding METR’s work, and it is in early talks with Hugging Face, CNBC reports.
For businesses choosing an AI supplier, Claude now has an outside safety checker whose findings Anthropic cannot veto, only redact in set areas.
Sources
- Anthropic: An update on independent evaluators. Checked
- TechCrunch: Anthropic is letting METR researchers look under the hood with employee-level access. Checked
- CNBC: Anthropic taps nonprofits METR and SecureBio to help assess risks and safety practices. Checked
Alexandre S.
Alexandre S. is a creative ad strategist and copywriter with ten years' experience in digital marketing across multiple channels.
More AI news from
- Meta releases Muse Spark Open, an AI model that works on Macs
- FTC sends legal demands to OpenAI and Anthropic over AI safety
- Amazon launches two Bee AI gadgets that listen and take notes
- Some Anthropic investors now eye a $1.5 trillion IPO valuation
- AI chip demand gives TSMC its sixth record quarter in a row
- Accused OpenAI engineer booked a one-way flight, prosecutors say
- Nvidia releases an open model built to run AI agents all day
- Flaws found by AI keep 1,000 contractors busy at Bank of America
- AI agent payment system x402 passes 138 million transactions