The United States government has introduced a new system to evaluate artificial intelligence models before they are made available to the public, focusing on potential risks linked to cybersecurity, biosecurity and chemical weapons.
Under the initiative, leading technology companies including Google, Microsoft and xAI have agreed to allow their AI systems to be assessed by the US Department of Commerce through the Center for AI Standards and Innovation (CAISI). The programme is designed to test advanced AI tools before release and identify potential safety and security concerns.
According to officials, the evaluations will include testing, collaborative research and the development of best practices for commercial AI systems. CAISI will also examine what it describes as “demonstrable risks” associated with frontier AI models, including threats related to cyberattacks and sensitive biological or chemical applications.
CAISI Director Chris Fall said independent assessment is essential to understanding the capabilities and risks of advanced AI systems.
“Independent, rigorous measurement science is essential to understanding frontier AI and its national security implications,” he said. “These expanded industry collaborations help us scale our work in the public interest at a critical moment.”
Microsoft confirmed that its participation in the programme will help strengthen protections against emerging threats, including AI-driven cyberattacks, particularly within its Copilot system.
The move represents a shift in approach for US AI policy under President Donald Trump, who has previously emphasised reducing regulatory barriers in order to maintain American leadership in artificial intelligence and prevent competitors such as China from gaining an advantage.
Earlier this year, the administration’s AI National Policy Framework called for accelerating AI deployment across industries while avoiding the creation of new federal agencies specifically dedicated to AI regulation. Instead, it placed responsibility on existing regulators and technical experts to oversee safety and compliance.
Despite this deregulatory stance, CAISI has already carried out around 40 evaluations of AI models, including systems that have not yet been publicly released. Officials have not disclosed details of those models.
The current agreements build on earlier arrangements made in 2024 with companies such as OpenAI and Anthropic under the previous administration. CAISI has confirmed that those agreements have now been renegotiated, although specific changes were not detailed.
OpenAI chief global affairs officer Chris Lehane said the company provided early access to its ChatGPT5.5 model ahead of its public release to assist with government-led security testing and evaluation.
Industry observers say the programme reflects growing concern over the dual-use nature of advanced AI systems, which can offer significant economic and scientific benefits while also posing risks if misused.
CAISI said the initiative is intended to strengthen cooperation between government and industry as AI systems become increasingly powerful and widely deployed.
