US Holds Off on Widening Export Controls for Anthropic AI Models

The U.S. government has held off on expanding export controls on artificial intelligence (AI) company Anthropic's latest models. The U.S. government had previously issued export control directives limiting foreign organizations and individuals' access to Anthropic's 'Mythos5' and 'Fable5' models, and Anthropic subsequently blocked all user access to comply with these directives.

David Saxe, co-chair of the President's Council of Advisors on Science and Technology, stated that a trusted partner had found a way to bypass safety features. The 'trusted partner' mentioned by Saxe turned out to be Amazon. Amazon CEO Andy Jassy reported that their researchers discovered a vulnerability in the 'Fable5' model that allowed safety features to be circumvented through specific prompts. U.S. government officials verified Amazon's claims, concluding that blocking access by foreign governments, corporations, and individuals was the most direct way to mitigate the risk.

Anthropic countered that the possibility of bypassing safety features was exaggerated and that similar levels of circumvention were possible with other models, including those from OpenAI. However, the U.S. government maintained its stance that such possibilities could not be classified as 'not serious'.

Amazon, a key investor in Anthropic since 2023, has invested approximately $13 billion in the company and has pledged further investment. Amazon stated, 'It is not uncommon for the government to seek advice on potential security risks.'

The measure received final approval from President Donald Trump, and a senior White House official explained that the action was related to model safety. Saxe argued that linking the measure to a conflict between the Department of Defense and Anthropic would be misleading.