US adds Google DeepMind, Microsoft and xAI to CAISI testing
The US is stepping up its AI safety game by adding Google DeepMind, Microsoft, and xAI to a government program that tests unreleased AI models for major risks.
Led by the Center for AI Standards and Innovation (CAISI), the program looks for issues like hacking threats, misuse for dangerous weapons, and corrupted training data, basically making sure these powerful AIs don't get out of hand.
OpenAI allows US testing of GPT-5.5-Cyber
OpenAI is letting the government test its latest GPT-5.5-Cyber model designed to help defend against cyberattacks.
Microsoft is building shared datasets and workflows (but keeping its specific models under wraps), while Google DeepMind has agreed to open up access to its own models.
Anthropic even let researchers try to break into their systems, and fixed any loopholes they found.
This all builds on past findings where flaws in OpenAI's ChatGPT Agent showed just how important these checks are for keeping AI safe.