Category: AI / Policy / Security



Every major U.S. frontier AI lab is inside the government's tent. Bloomberg reports that with Google, Microsoft, and xAI signing new pre-release evaluation agreements with CAISI, OpenAI and Anthropic — which had been in the program since 2024 — renegotiated their own deals to get in lockstep with Trump's AI Action Plan. The result? A sweep. Not a single major U.S. AI lab is sitting on the sidelines anymore.

OpenAI moved fast to underscore the point. Policy chief Chris Lehane confirmed in a LinkedIn post that the company had already handed the government early access to GPT-5.5 for national security testing — signaling that OpenAI isn't just playing along, it's getting out ahead of the pack. Anthropic, meanwhile, quietly updated its own arrangement — a notable move given that the company is simultaneously neck-deep in a separate and very public fight with Washington.

That fight is worth spelling out. The Pentagon designated Anthropic a supply chain risk in March after the company refused to strip guardrails on autonomous weapons — a move a federal judge later called "Orwellian." Defense Secretary Hegseth and President Trump have both laid out a six-month phaseout of Anthropic tools in government use, and two active lawsuits remain unresolved. Yet Anthropic's CEO Dario Amodei still walked into the White House just days after Mythos was unveiled, sat down with senior Trump officials, and both sides called it "productive." Welcome to the most complicated relationship in Washington.

CAISI itself is operating on shaky institutional ground. The center — originally the Biden-era AI Safety Institute, renamed last June by Commerce Secretary Howard Lutnick who called the old name a cover for regulation — still lacks permanent legal standing. Its director, Chris Fall, stepped into the role after Collin Burns was pushed out just four days in over concerns about his Anthropic ties. Some lawmakers are pushing legislation to codify CAISI, but nothing has crossed the finish line.

And here's where the stakes go global. With nearly every top U.S. AI lab now locked into a common pre-deployment evaluation regime, Washington is quietly setting the template for how the rest of the world handles frontier model oversight. The UK's AI Security Institute already runs comparable assessments. The EU just softened its AI Act. Three separate evaluation regimes — CAISI, AISI, and a trimmed-down EU AI Act — are clearly not converging, leaving companies operating across all three jurisdictions to navigate a patchwork of overlapping rules. That's not just a compliance headache; it's a preview of the global AI governance battle that's just getting started.

Official source:
U.S. Commerce Department / Bloomberg

Original official source URL:
https://www.commerce.gov/

Source article URL:
https://www.bloomberg.com/news/articles/2026-05-05/ai-firms-agree-to-give-us-early-access-to-evaluate-their-models