Anthropic Brings in External Safety Oversight Team, Becoming First Frontier AI Company to Deliver on Promise
nashnova research
Anthropic has embedded six independent researchers inside the company with near-employee-level access to audit unreleased AI models — the first time a frontier AI firm has moved an external safety oversight pledge from paper to practice, a step rivals promised but have yet to take.
What actually happened?
Anthropic brought in six evaluators from independent research group METR, granting them "employee-like" access to company hardware, office space, and internal tools.
Their job: conduct deeper safety assessments of AI models and agents before public release.
This means → external oversight is no longer limited to reading published papers; these researchers can touch the same systems insiders use.
Why is this called a first?
Last month, Anthropic CEO Dario Amodei published an essay urging frontier AI companies to slow down and unilaterally pledged to embed external safety evaluators.
OpenAI CEO Sam Altman responded that his company would do something similar — but has not yet followed through.
In plain terms = several companies have said "we'll do safety oversight." Anthropic is the first to actually put people inside the building and hand them the keys.
What can these evaluators say — and what can't they?
Anthropic lets METR publish its findings freely, with three carve-outs: no disclosure of customer data, model architecture, or attorney-client privileged information.
Anthropic can ask METR to redact content — but METR may publicly state that redactions were made and whether it considers them justified, as long as it does not reveal the redacted material itself.
This means → external monitors hold real power, yet commercial secrets remain fenced off. Whether the mechanism is robust enough will only become clear once the first evaluation reports are published.
Who is METR, and why them?
METR is an independent AI-safety research organization, known in the industry for publishing a report on security vulnerabilities in OpenAI's agents.
The six selected researchers have not been publicly named and will be firewalled from METR's broader team to meet Anthropic's security standards.
This reflects a practical reality: organizations capable of evaluating frontier AI safety are very few. METR has both the expertise and the credibility, but the firewall shows both sides are wary of information leakage.
市场有风险，内容仅供研究参考，不构成投资建议。