The Trump administration, which once promised Silicon Valley it would never slow down artificial intelligence with “foolish rules,” is now drafting legislation that could require tech companies to submit their most powerful AI models for government vetting before public release. The policy reversal, first reported by the New York Times and confirmed by Axios, marks one of the sharpest about-turns in recent tech policy history: an administration that scrapped Biden-era AI safety rules on its first day in office is now considering rebuilding a version of exactly those rules, just with different agencies in charge.The catalyst is Anthropic’s Mythos, an AI model so capable of autonomous hacking that the company itself refused to release it to the public. Since its limited unveiling last month, Mythos has alarmed national security officials, rattled global banks, forced a reckoning inside the White House, and quietly changed what Washington thinks AI regulation should actually look like.
From ‘beautiful baby’ to bureaucratic oversight, what changed Trump’s mind on AI regulation
Mythos Preview is currently available only to roughly 40 handpicked tech and cybersecurity companies, and once you understand what it can do, the administration’s reversal starts to make sense.The model doesn’t just find software vulnerabilities. It hunts them down autonomously, at a scale no human researcher can match. In internal testing, Mythos uncovered thousands of critical zero-day flaws across every major operating system and web browser. It found a 27-year-old vulnerability in OpenBSD, one of the most security-hardened systems in the world. It cracked Linux kernel flaws and chained them together to allow complete machine takeover. It exposed a 16-year-old bug in FFmpeg—a piece of software used by countless applications to process video—that automated tools had scanned five million times without catching.Anthropic’s own red team lead, Logan Graham, told Axios the model can find “tens of thousands of vulnerabilities” and write working exploits to go with them. Its predecessor, Claude Opus 4.6, found roughly 500 zero-days. Mythos found multiples of that, faster, with less human input, and with a success rate of 83.1 percent on reproducing known vulnerabilities on the first attempt.Washington noticed quickly. The NSA quietly used Mythos to probe weaknesses in US government software. Agencies that had been cut off from Anthropic’s tools following a bitter Pentagon contract dispute suddenly wanted back in. And the White House, which had spent months trying to freeze Anthropic out of government contracts entirely, found itself in the awkward position of needing the company it was simultaneously fighting in court.
The proposed AI safety review law would give government early access to models, not a veto over them
The executive order now under discussion would establish a formal AI working group made up of tech executives and senior government officials. According to the New York Times, the group would develop pre-release oversight procedures for frontier AI models, potentially modelled on the UK’s approach, where multiple government bodies evaluate AI systems against safety standards before and after deployment.The proposed framework would not give the government power to block a model’s release outright. It would grant early access to agencies like the NSA, the Office of the National Cyber Director, and the Director of National Intelligence. Axios reported that the ONCD has already hosted two rounds of meetings with tech companies and trade groups to sketch out a framework requiring the Pentagon to lead safety testing for AI deployed across federal, state, and local governments.A White House official told the Times that any talk of an executive order is “speculation” and that announcements will come directly from the president. But Axios noted the underlying security framework is “fairly far along” and was already in development before Mythos made the conversation feel urgent.