Anthropic announced in the early hours of October 16th (Beijing time) that it has granted six researchers at the non-profit AI assessment organization METR near-employee-level internal access, including access to company equipment, office space, and internal tools, enabling them to conduct in-depth assessments of the security risks of cutting-edge AI models. METR can independently publish assessment reports on model risks, anomalies, and the company's security practices. Anthropic cannot veto the publication of these reports, but it can request the removal of pre-defined categories of information, such as customer data, trade secrets, and model architecture. METR has the right to publicly evaluate the appropriateness of these deletions and modifications. Anthropic also proposed a three-tiered independent assessment mechanism, with the highest level allowing assessors in-depth access to the company and independent publication of conclusions. Currently, SecureBio is at the first tier, responsible for reviewing risks related to biological weapons; Accenture and its Faculty are also involved in model security assessment collaborations. Anthropic stated that it hopes other leading AI companies will adopt similar measures to improve the transparency and credibility of industry security oversight.
