US presses Meta to submit AI models for government safety review

The US government is pressing Meta to submit its AI models for official safety evaluations, The New York Times reported Monday (local time), as Washington moves to build a formal AI safety verification framework.

Citing multiple anonymous sources, the Times said the government has exchanged confidential emails with Meta and is pushing the company to sign an AI safety assessment agreement.

Meta is the only major US AI developer that has not joined the government's safety evaluation framework.

OpenAI, Anthropic, Google, xAI and Microsoft have already submitted their latest models to the Center for AI Safety and Innovation, or CAISI, the government body responsible for AI safety assessments.

Meta unveiled its latest AI model, Muse Spark, in April. Industry observers consider the model broadly competitive with the latest offerings from its major rivals.

Meta spokesperson Frances Brennan said the company shares the government's goal of securing American leadership in powerful and safe frontier AI. "We are currently working through the specific details and hope to be able to sign an agreement soon," she said.

Ben Kass, a Commerce Department spokesperson, said CAISI regularly engages with companies on such agreements. "This is standard practice — it's exactly what CAISI does," he said.

About two weeks ago, the government ordered that foreign access to Anthropic's latest models, Fable 5 and Mythos 5, be blocked on national security grounds.

The government and Anthropic subsequently held talks over the matter. President Donald Trump said in a recent interview with Axios that he no longer views Anthropic as a national security threat.

Trump established the government's safety review process on June 2, when he signed an executive order requiring the government to assess the security of new AI models.

Under the order, technology companies must allow the government up to 30 days to evaluate a new AI model before its public release. During that period, the government can assess the model's capabilities and vulnerabilities.