OpenAI and Anthropic Embed Independent AI Safety Evaluators
AI labs OpenAI and Anthropic are integrating third-party evaluators to assess model safety following a voluntary accord presented by President Donald Trump.
Leading AI laboratories OpenAI and Anthropic are integrating independent third-party evaluators to assess model capabilities and safety risks. This shift follows a voluntary accord presented by President Donald Trump in late September, which encouraged AI companies to partner with external auditors. Anthropic is currently embedding employees from Accenture to test safeguards and assess model behavior, a move endorsed by OpenAI CEO Sam Altman.
While the labs are partnering with organizations such as Model Evaluation and Threat Research (METR) and Apollo Research, internal friction has emerged. OpenAI recently fired three employees, including Mikita Balesni and Tomek Korbak, who claim their dismissals were linked to communications with these third-party evaluators.
In the absence of federal regulation, California Governor Gavin Newsom signed legislation establishing a state registry for AI auditors and a framework for Independent Verification Organizations (IVOs). At the federal level, Congress has seen the introduction of the FRONTIER Act to create a national IVO system. Critics and academics warn that the current ecosystem lacks robust standards and financial independence, potentially leaving evaluators beholden to the labs they are tasked with auditing.
