Anthropic Grants Outside Evaluator METR Employee-Level Access in AI Safety Push

3-line summary

• Anthropic announced in an October 15 blog post (reported by Reuters the same day) that it has granted outside AI-safety evaluator METR employee-level access — six METR staff can access Anthropic's hardware, office space, and internal systems at a level roughly equivalent to Anthropic's own risk-assessment team, excluding customer data, legally privileged information, and model architecture.
• Under the arrangement, METR can publish findings about Anthropic's risks, incidents, and practices without a veto, with Anthropic limited to narrowly defined redactions that METR can publicly challenge without having to reveal what was redacted; this is the third of three access tiers Anthropic is developing, with METR's publication rights similar to tier two. Other partners include SecureBio (reviewing the chemical/biological weapons chapter of an August risk report), Faculty (Accenture's AI unit, announced in September for model evaluation and adversarial testing), and Hugging Face (in discussions).
• The move follows a Reuters report earlier the same week that Anthropic and OpenAI sought significant control over what outside reviewers could publish, citing trade-secret concerns; on October 14, METR had proposed that developers allow confidential staff interviews, timely disclosure of evaluation results, and transparency about access limits and redactions.

This summary is our own processing of the source article; see the source link below for full context.