“Anthropic has granted METR researchers employee-level access to evaluate its frontier AI models, risks, and internal practices. Critically, Anthropic has waived its right to block or veto METR's published findings. This marks a significant step toward independent third-party oversight of a leading AI lab.”
Key Takeaways
- METR receives employee-level internal access to Anthropic's systems, models, and practices — not just curated demos.
- Anthropic has contractually surrendered the right to block or suppress METR's published assessments.
- METR will independently publish evaluations covering Anthropic's risks, incidents, and safety practices.
METR gains unprecedented employee-level access to audit Anthropic's AI risks independently.
trending_upWhy It Matters
This arrangement sets a rare precedent in the AI industry, where labs have historically controlled the narrative around their own safety claims. By ceding veto power over published findings, Anthropic is accepting genuine external accountability rather than performative audits. If this model is adopted more widely, it could become the template for meaningful third-party AI oversight — something regulators in the EU and US have been struggling to mandate. Competitors like OpenAI and Google DeepMind will face pressure to offer similar access, or explain why they won't.
FAQ
Who is METR and why are they qualified to audit Anthropic?
METR (Model Evaluation and Threat Research) is an independent AI safety organisation focused on evaluating the risks posed by frontier AI models. They specialise in assessing dangerous capability thresholds and have been involved in pre-deployment evaluations for several leading AI labs.
Can Anthropic interfere with or suppress METR's findings?
No — Anthropic has explicitly given up the right to block METR's publications as part of this arrangement. This means METR can publish critical findings about risks or incidents without Anthropic's approval.
Does this mean Anthropic's models have been found to be unsafe?
Not necessarily — the agreement is about establishing ongoing independent oversight, not a response to a specific safety failure. It reflects a proactive commitment to transparency, though METR's forthcoming assessments will reveal what they actually find.
