Washington is writing AI regulation with Anthropic, OpenAI and Google at the table and calls that independent oversight

Anthropic claims to have built the most dangerous AI system in existence right now, but whether that’s true or just a good marketing pitch doesn’t matter, because the company then invited the government to the table, there was a federal contract still open so they could discuss that too, and is now co-writing the law that determines whether something like this is allowed to happen again. An urgency that is entirely managed by the party that created it isn’t an urgency with a constructed narrative around it; it is the construction. Anthropic made the danger, named it, and controlled access to it through forty hand-selected companies. The danger is the product, the regulation is distribution control, and the government contract covers the development costs.

How AI safety works when Anthropic decides who gets access to the danger Anthropic created

Project Glasswing: forty companies, exclusive access to the vulnerability intelligence Mythos exposed. Google on the list, Microsoft on the list, xAI on the list; the central bank of Mozambique, not so much. The parties that need that information most have it, and Anthropic calls this responsible stewardship. The FDA framework assumes you know what you’re testing for; Mythos is emergent, nobody fully understands what it can do until it does it. A framework that structurally cannot evaluate what it claims to evaluate still protects the company that cooperated, signed off, and “demonstrated responsibility.” The EPA regulated lead pollution for years using standards the lead industry helped write. That’s not a precedent. That’s an instruction manual.

How AI safety became a market entry barrier and why that was always the point

Three months ago the Vice President called AI safety concerns fear; now “safety” appears three times in a single White House post. That’s not a policy pivot, it’s a vocabulary swap by people who know that words are cheaper than oversight. Safety now means what Anthropic, OpenAI and Google mean when they say it: their standards, their definitions, their solutions. Smaller labs will soon navigate an approval process built on standards written by the companies they compete with; safety requirements as market entry barriers are as old as the pharmaceutical industry. The structure meant to handle the moment when a frontier lab wants to release something the government considers dangerous already exists, and it was built by the people who will cause that moment.