US Secures Access to Top AI Models for Security Scrutiny

The US federal government has secured agreements with Micrososources, Google, and xAI to access their advanced artificial intelligence models for national security testing. This initiative aims to evaluate AI capabilities and identify potential security risks before widespread deployment.

The Center for AI Standards and Innovation (CAISI) will conduct these evaluations. The focus is on probing unexpected behaviors and developing shared datasets for testing the models.

This development follows increasing concerns regarding the potential for advanced AI systems to be exploited for malicious purposes, including cyberattacks. The agreements build upon previous collaborations with other AI developers.

CAISI has already conducted over 40 evaluations, including on models not yet publicly available. Developers are providing versions of their models with safety guardrails removed to facilitate comprehensive risk assessment.

This move aligns with a broader government strategy to vet AI models for national security implications. It also follows a separate Pentagon agreement with seven tech companies to integrate AI into classified systems.