Trump signs executive order on AI to encourage top AI models to be tested before release.
(Washington, D.C.) U.S. President Donald Trump signed an executive order on artificial intelligence (AI) to establish a voluntary framework that allows AI developers to open their models to the federal government for cybersecurity testing before publicly releasing their most advanced models.
This order applies to leading AI companies such as OpenAI, Google, and Anthropic. The 30-day testing window is a compromise reached after a tug-of-war between the White House and the tech industry; an earlier draft proposed a maximum of 90 days for government pre-review, while tech companies pushed to shorten the period to 14 days.
This executive order was driven by security concerns raised by Mythos, an AI model developed by Anthropic. Anthropic had previously kept the model private because of its ability to discover vulnerabilities in computer systems, including those in critical sectors such as banks, government agencies, and hospitals.
To address the associated risks, the executive order requires federal agencies to collaborate with AI companies and critical infrastructure operators to establish an AI cybersecurity information-sharing mechanism to identify software vulnerabilities and promote patching efforts.
This policy also reflects a shift in Trump's approach to AI governance. Since returning to the White House, Trump has advocated for a less interventionist approach to federal regulation of the tech industry and has attempted to dissuade states from implementing AI regulations he opposes. However, as concerns in Washington have grown regarding the cybersecurity risks of advanced AI systems, the White House has begun seeking to allow the government earlier access to high-risk models without mandatory approval mechanisms.
Trump was originally scheduled to sign the executive order on May 21, but postponed it at the last minute. He stated at the time that he was dissatisfied with parts of the order and did not want to take any measures that could weaken the United States' position in the AI competition with China. The final version shortened the previously discussed 90-day review window to a maximum of 30 days and emphasized voluntary participation by companies.
AI security experts believe that the voluntary framework is still insufficient to address the national security threats that high-risk models may pose. How governments will define which models need to be tested in advance, and whether companies will continue to cooperate, remains to be seen.
Source: [Lianhe Zaobao] (https://www.zaobao.com/news/world/story20260603-9149332)