US Expands Pre-Deployment Vetting of Google, xAI, Microsoft AI Models

The CAISI (Center for AI Standards and Innovation), under the US Department of Commerce, has expanded its system for evaluating unreleased AI models before launch to include Google DeepMind, xAI, and Microsoft.

The pre-deployment vetting system, in which OpenAI and Anthropic already participated, has now broadened to major US frontier AI labs, with the goal of catching risks such as cyberattacks, military misuse, chemical and biological weapons development, and training data contamination in advance.

• OpenAI is testing GPT-5.5-Cyber with CAISI.
• Microsoft is building shared datasets and test workflows with government scientists.
• Google DeepMind is providing proprietary models and data.
• CAISI has already completed more than 40 evaluations, including of models not yet made public.

Developers often hand over versions with reduced safeguards so that CAISI can test attack scenarios, and in this process red-teaming has revealed bypass techniques and cyber vulnerabilities. CAISI is also preparing testing guidelines for AI used in critical infrastructure such as telecommunications and emergency services.