Key Points
- Google DeepMind, Microsoft and xAI join OpenAI and Anthropic in pre-release AI model reviews
- US government body CAISI has completed over 40 evaluations including unreleased models
- Agreements allow government assessment of AI capabilities before public deployment
Google DeepMind, Microsoft and xAI have agreed to give the United States government early access to their artificial intelligence models before public release, allowing federal evaluators to assess the systems’ capabilities and security risks.
The three companies join OpenAI and Anthropic in permitting pre-deployment reviews by the Center for AI Standards and Innovation (CAISI), a body under the US Commerce Department that conducts safety evaluations of frontier AI systems — the most advanced AI models being developed by leading laboratories worldwide.
The agreements allow CAISI to evaluate AI models before they reach the public, assess systems after deployment and conduct targeted research into AI security. The body said it has completed over 40 such evaluations to date, including assessments of state-of-the-art models that remain unreleased.
What the expanded partnerships mean
Pre-deployment evaluation means government researchers can test an AI system’s capabilities, identify potential misuse scenarios and flag security vulnerabilities before the model becomes publicly available. This approach gives regulators visibility into what the most powerful AI systems can do before millions of users gain access.
CAISI, formerly known as the Artificial Intelligence Safety Institute, first announced agreements with OpenAI and Anthropic in August 2024. The body said the new agreements with Google DeepMind, Microsoft and xAI build on those earlier partnerships, which have been renegotiated to align with directives from the secretary of commerce and the administration’s AI Action Plan.
“Independent, rigorous measurement science is essential to understanding frontier AI and its national security implications,” said Chris Fall, director, CAISI. “These expanded industry collaborations help us scale our work in the public interest at a critical moment.”
Defence deployment agreements
The CAISI agreements follow a separate development last week, when the Department of War signed contracts with several technology companies to deploy AI capabilities on classified military networks.
Those agreements, distinct from the CAISI evaluation partnerships, involve OpenAI, Google, Microsoft, Amazon Web Services, Nvidia, Oracle and Reflection AI. The defence contracts authorise deployment of advanced AI systems for lawful operational use within the department’s classified infrastructure.
The expansion of government access to unreleased AI models reflects growing concern in Washington about the national security implications of advanced AI systems.
Frontier models developed by leading laboratories can generate text, code, images and video with increasing sophistication, raising questions about potential misuse for disinformation, cyberattacks or the development of dangerous capabilities.
For Indian technology companies and researchers tracking global AI governance, the US approach of pre-deployment government evaluation represents one model for how regulators might seek visibility into advanced AI systems before public release.
A comparable logic has already been applied domestically in the telecommunications sector, where India’s National Security Directive mandates that operators procure equipment only from government-designated “trusted sources,” with vendors and products subject to prior security clearance before deployment.
This precedent suggests that Indian regulators are not unfamiliar with ex-ante oversight of critical technologies, and similar frameworks could plausibly be extended to frontier AI systems.
Your Questions, Answered
What is CAISI and what does it do?
The Center for AI Standards and Innovation is a US Commerce Department body that evaluates frontier AI models for security risks and capabilities. It conducts assessments before and after AI systems are released to the public.
Which companies have agreed to pre-deployment AI evaluation?
Google DeepMind, Microsoft and xAI have now joined OpenAI and Anthropic in allowing CAISI to evaluate their AI models before public release.
What is frontier AI?
Frontier AI refers to the most advanced artificial intelligence models being developed by leading laboratories. These systems represent the cutting edge of AI capabilities in areas such as text generation, coding and reasoning.
How many AI evaluations has CAISI completed?
CAISI says it has completed over 40 evaluations to date, including assessments of state-of-the-art models that have not yet been released to the public.
