On October 15, 2026, Anthropic introduced a new three-tier evaluation framework for its AI models. The initiative grants independent researchers deeper access to the company's internal systems and hardware.
A six-member team from the nonprofit METR has been granted employee-like access to Anthropic's hardware and internal systems. METR holds the right to publish its findings without prior approval from the company, subject only to limited redactions for sensitive information.
The framework assigns specific roles to external partners. The nonprofit SecureBio operates at tier one, focusing on biosecurity and reviewing risk reports, including the August 2026 report on chemical and biological weapons. Previously, SecureBio assessed the Claude Opus 4.6 model.
Accenture’s Faculty unit manages tier two, which covers adversarial testing and broader evaluations. Anthropic is also reportedly considering a potential collaboration with Hugging Face to further expand its network of safety reviewers, though no partnership has been confirmed.
