SAN FRANCISCO — OpenAI on Thursday published expanded details of its strengthened safety framework, outlining new monitoring, alignment, and misuse-detection measures for its most capable AI models. The company said the changes follow a policy review announced earlier in the week.

The move builds on a blog post OpenAI released Tuesday, in which the company said it was tightening internal oversight of how its models are trained and deployed. OpenAI has faced sustained pressure from researchers and regulators over risks posed by advanced systems, including potential misuse in cyber operations and biological research.

The updated framework introduces automated monitoring of model outputs, tiered access controls for higher-risk capabilities, and expanded red-teaming before major releases. OpenAI said it would also strengthen collaboration with external safety evaluators and government bodies, including the US AI Safety Institute.

The announcement comes as governments in the United States and European Union advance rules requiring greater transparency from frontier AI developers. Under the EU AI Act, providers of general-purpose models face obligations on documentation, risk assessment, and incident reporting, with enforcement stages continuing through 2026.

Safety researchers have argued that voluntary commitments from AI firms require independent verification. OpenAI said it would publish periodic reports on its safety performance, though it did not specify how external parties would audit the results.