The U.S. government has reached agreements with leading AI developers Google DeepMind, Microsoft, and xAI to facilitate the review of early versions of their advanced AI models prior to public release. These evaluations, managed by the Center for AI Standards and Innovation (CAISI), part of the U.S. Department of Commerce, are deemed crucial for understanding the capabilities of powerful new AI models and for safeguarding U.S. national security. These collaborations are intended to help the federal government scale its work in the public interest during a pivotal period for AI.
Chris Fall, CAISI director, underscored that "independent, rigorous measurement science is essential to understanding frontier AI and its national security implications." CAISI's mandate is to facilitate collaboration between the tech industry and the federal government to develop standards and assess risks for commercial AI systems. The current agreements specifically target national security risks linked to cybersecurity, biosecurity, and chemical weapons.
These new partnerships expand upon similar deals that the Biden administration forged with OpenAI and Anthropic two years ago. CAISI has already successfully completed over 40 such evaluations, including assessments of unreleased models. It is a recognized practice for developers to share unreleased AI models with the government, often with reduced or even removed safety guardrails. This intentional approach allows the government to conduct a more thorough evaluation of national security-related capabilities and potential risks inherent in these cutting-edge systems.
The increased focus on pre-release scrutiny arises from escalating fears that the newest and most powerful AI models, such as Anthropic’s Mythos, could pose significant public dangers. AI safety experts, government officials, and tech companies are concerned that the expansive capabilities of these models could enable hackers to exploit cybersecurity vulnerabilities on an unprecedented scale. In response, Anthropic limited the initial rollout of Mythos to a select few companies and initiated Project Glasswing, a collaborative industry effort aimed at securing the world's most critical software.
Echoing these efforts, Microsoft also announced a similar agreement in the UK with the government-backed AI Security Institute, which shares a focus on safe AI development. Microsoft articulated in a blog post that while it regularly undertakes various forms of AI testing internally, assessing national security and large-scale public safety risks necessarily requires a collaborative endeavor involving governments.