On Tuesday, Trump signed a new executive order on AI, requiring federal agencies to leverage advanced artificial intelligence to enhance cybersecurity capabilities and expand collaboration with major AI companies. The White House positioned the arrangement as a national security measure, while emphasizing that the release of new models would not become a formal approval process.
Establish a frontier model review mechanism
Under the executive order, the U.S. government will establish a confidentiality review process, with the National Security Agency determining which advanced AI systems fall under the category of "cutting-edge models." Developers can voluntarily submit their models for government evaluation up to 30 days before sharing them with other trusted partners.
The key focus of this mechanism is to identify whether a model possesses capabilities that could impact cybersecurity or national security before it is released on a larger scale. The executive order also requires coordinated action among federal departments to expedite the deployment of technologies deemed secure.
Previously postponed due to concerns about competition with China
Prior to this executive order, Trump had delayed signing a similar plan in May. According to reports, he believed at the time that some of its provisions could slow down U.S. AI development and weaken the U.S.'s leading position in competition with China.
The officially signed version still attempts to maintain a balance between security review and industry development. The document explicitly states that the new framework should not be interpreted as a mandatory licensing system before model release, which is also a response to the concerns of AI developers.
The controversy centers on voluntary submission for review.
Critics argue that the arrangement relies too heavily on voluntary corporate cooperation and lacks sufficient binding force. Policymakers at the U.S. consumer advocacy group Public Citizen state that if the model has become powerful enough to potentially threaten cybersecurity and national security, more enforceable federal regulations are needed, including transparency requirements, independent testing, and enforceable safeguards.
One of the factors pushing the White House to expedite the drafting of the executive order was the concerns raised by Anthropic's Claude Mythos model. This model demonstrated its ability to identify software vulnerabilities, prompting U.S. officials to reassess the security risks posed by highly capable AI models.
The report mentioned that in April of this year, after Mythos's capabilities were revealed, U.S. Treasury Secretary Bessant and then-Federal Reserve Chairman Powell convened a meeting with executives from major Wall Street banks to discuss the cybersecurity risks that the new AI model might pose.
The scope of law enforcement has also been included.
The executive order also calls for a tougher crackdown on the criminal use of AI, including hacking into public or private information systems or using AI agents to illegally obtain data and use it for criminal activities.
Last month, U.S. federal prosecutors indicted two men, accusing them of using AI to generate and distribute explicit images of women without their consent. This is seen as one of the earlier and most significant enforcement actions under the new legal framework.
Additional information:Anthropic also announced on the same day that it is expanding the limited access to Claude Mythos through Project Glasswing, allowing technology companies, security agencies, and governments to test potential vulnerabilities before the model is publicly released. The company had hinted last week that the model would be further rolled out in the coming weeks.
