# The Trump administration has expanded a program to give U.S. government scientists access to unreleased artificial intelligence models.

*genai · news · 2026-05-05 · Livemint*

## Key points

- The U.S. has expanded AI risk assessment access to DeepMind, xAI, and Microsoft models.
- OpenAI is sharing a new variant, GPT-5.5-Cyber, specifically for defensive cybersecurity testing.
- Anthropic's collaboration revealed safety overrides like false human review and character substitution exploits.
- OpenAI identified a ChatGPT Agent flaw allowing attackers to impersonate users across websites remotely.
- U.S. government scientists are drafting AI testing guidelines for critical infrastructure sectors.

**Companies:** Microsoft, Google DeepMind, OpenAI, Anthropic, xAI
**Countries:** United States

[Read the full story on Livemint](https://www.livemint.com/technology/factboxwhat-we-know-about-us-stress-tests-of-google-xai-and-microsoft-ai-models-11778018927864.html)

---

Canonical: https://newsio.io/n/642f617f-e59c-45d0-8057-3ee3eb29dc42/the-trump-administration-has-expanded-a-program-to-give-u-s-government-scientist
Summarized by Newsio from Livemint. https://newsio.io/how-it-works