Anthropic has granted six staff members from the non-profit organization METR access rights nearly equivalent to those of its internal employees for large models. These staff members can be stationed at Anthropic's offices, accessing its hardware and internal systems to evaluate the risks associated with large models. The evaluation reports will be independently published by the assessors, with Anthropic only able to request the redaction of specific information such as customer data, confidential content, and model architecture details, but without the authority to veto the report's release. This collaboration marks Anthropic's first initiative to fulfill its previous commitment to station external auditors to strengthen AI safety protections. METR operates with independent funding and does not charge Anthropic any fees. Currently, as concerns grow over the potential risks of highly capable AI systems, introducing third-party evaluations has become one of the core strategies employed by leading AI labs. Anthropic's approach has received joint endorsement from Sam Altman of OpenAI and Elon Musk of SpaceX, although OpenAI has yet to announce a related implementation plan. However, this mechanism has also sparked criticism over conflicts of interest, with calls for stricter government regulation. Anthropic argues that complete information disclosure could be maliciously exploited, likening its model to the mechanism where banking regulatory authorities station personnel within large banks to review risks. Additionally, Anthropic is advancing plans for an IPO within the year, with an expected valuation exceeding $2 trillion, and has highlighted its safety control measures in investor presentation materials. Besides METR, Anthropic is also collaborating with Accenture, investing at least $1 billion to establish a safety protection system, and is jointly reviewing risks related to chemical and biological weapons with the third-party organization SecureBio. Anthropic calls on all frontier AI companies to adopt the practice of stationing external evaluators.