人工智慧公司Anthropic正式向非营利评估机构METR授予「如同员工般」的内部权限，以全面评估其AI模型的潜在风险。这项举措落实了执行长Dario Amodei先前的安全承诺，旨在应对强大AI系统可能失控并对经济、网路安全及人类社会造成灾难性影响的担忧，同时这也是去年九月OpenAI与SpaceX高层共同支持引入外部审计提议后，首个付诸实行的具体步骤。
在具体执行层面，Anthropic已让六名METR人员进驻办公室，提供公司硬体与内部系统权限，并鼓励员工向其分享风险资讯。为避免利益冲突，Anthropic不会向METR支付费用，而METR也不接受AI企业的捐款，其资金主要来自慈善机构与英国AI安全研究所等政府单位。此外，Anthropic无权否决METR发布的评估报告，仅能要求隐去客户资料、法律特权细节或模型架构等敏感商业机密。
面对外界对审计独立性的质疑，以及估值逾两兆美元的首次公开募股（IPO）所带来的投资人压力，Anthropic将此机制比拟为银行业的驻场监管制度，强调能在保护商业机密的同时实现公众问责。除了METR之外，Anthropic也持续与埃森哲（Accenture）共同投入至少十亿美元建立新防护措施，并与SecureBio合作审查生化武器风险，同时呼吁所有前沿AI开发商都应自愿引入外部评估人员。（关键数字：2, 1）
Artificial intelligence company Anthropic has officially granted "employee-like" internal access to the non-profit evaluation organization METR to thoroughly assess the potential risks of its AI models. This move fulfills a prior safety commitment by CEO Dario Amodei to address growing concerns that powerful AI systems could run catastrophically out of control and threaten the economy, cybersecurity, and humanity at large, marking the first concrete implementation of a proposal to embed external auditors that was also backed by OpenAI and SpaceX leaders in September.
In terms of practical execution, Anthropic has onboarded six METR staff members into its offices, providing them with company hardware and internal system access roughly equivalent to its own risk teams, while encouraging employees to openly share information with them. To prevent conflicts of interest, Anthropic is not paying METR for this arrangement, and METR does not accept donations from AI companies, relying instead on funding from philanthropies and government bodies like the UK's AI Security Institute. Furthermore, Anthropic cannot veto the publication of METR's analysis, though it can require redactions for sensitive categories such as customer data, attorney-client privileged details, or model architectures.
Facing external skepticism regarding auditing independence and pressure to reassure investors ahead of an expected IPO valued at over $2 trillion, Anthropic likened this arrangement to resident bank supervision, emphasizing that it achieves public accountability without exposing sensitive information to malicious actors or competitors. Beyond METR, Anthropic continues to partner with Accenture in a joint commitment of at least $1 billion to establish new safeguards and works with SecureBio to review chemical and biological weapon risks, while calling on all frontier AI companies to voluntarily embed external evaluators.