US agency to review AI models before release

The US-American CAISI (Center for AI Standards and Innovation) has announced a new agreement with Google DeepMind, Microsoft, and xAI. As part of the collaboration, CAISI will evaluate the AI models of these providers before their release. The cooperation is focused on AI security and national security. In conjunction with the new agreement with Google, Microsoft, and xAI, CAISI has renegotiated an existing agreement with OpenAI and Anthropic.

The agreement between the AI providers and the agency includes, in addition to evaluation before the general availability of the models, post-deployment assessments and other unspecified research. This is said to have already happened with more than 40 models, some of which represent the state of the art and are not yet released. For its investigations, CAISI is to receive models with reduced or removed safety precautions. According to CAISI, the agreements with the tech companies promote information exchange and help the US government gain a clear understanding of AI capabilities and international competition.

What is CAISI?

CAISI is part of the NIST (National Institute of Standards and Technology), which publishes guidelines for software in government agencies or proposals for secure routers, for example. The agencies are under the United States Department of Commerce. CAISI's mission statement is to develop and publish voluntary guidelines that promote the responsible design, development, deployment, use, and governance of advanced AI models, systems, and agents. The center acts as the primary point of contact for the business community within the US government regarding AI. Microsoft has also announced a similar agreement with the UK's AI Security Institute.

CAISI's press release on the agreement with Microsoft, Google DeepMind, and xAI can be found here.

(pst)