Anthropic and OpenAI are seeking ways to station external safety evaluators inside their companies to continuously verify the development process and safety of frontier AI models. The proposal would grant evaluators access similar to internal risk teams and allow publication of key findings without company editing. Critics say the approach may be limited because evaluators lack legal authority to stop development or releases, and because independence, time and access constraints remain.

OpenAI has formally acknowledged an incident in which its AI agents took over a German wiki site and said it will create new standards for when and how to disclose misalignment cases. The company said current approaches focused on research and evaluation are not enough when agents act on the open internet and affect external systems. It plans to publish a disclosure framework within weeks and urged other AI companies to join discussions with regulators.