Trump administration considers mandatory pre-release vetting of AI models — Anthropic's Mythos cited as catalyst for policy reversal

News
By published

The company that wanted AI regulation under Biden may have gotten its wish, just not on its own terms.

Dario Amodei looking confused.
(Image credit: Michael M. Santiago via Getty Images)

The Trump administration is said to be discussing an executive order that would establish a government review process for new AI models before they’re released to the public, The New York Times has reported, citing unnamed U.S. officials.

The sudden reversal coincides with a leadership vacuum in White House AI policy. David Sacks, who led the administration's deregulation push as AI czar, left the role in March, with White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent having since taken a more active role in shaping AI policy, according to The New York Times.

Latest Videos From

The new approach sounds a lot like the UK's AI Security Institute model, where government bodies evaluate frontier models against safety benchmarks before and after deployment. Officials told the New York Times that the NSA, the Office of the National Cyber Director, and the Director of National Intelligence could oversee the review. Critically, the system would grant the government early access to models without blocking their release.

Perhaps unsurprisingly, the catalyst for all this appears to have been Anthropic’s Mythos model, which the company’s marketing described as capable of finding thousands of critical software vulnerabilities and too dangerous for public release.

That naturally attracted a lot of unwanted government attention at a time when the Trump administration is already locking horns with Anthropic over the collapsed $200 million Pentagon contract. The Pentagon designated Anthropic a supply chain risk after the company refused to remove guardrails on autonomous weapons and mass surveillance, though a federal judge later called that "Orwellian."

The NSA has already used Mythos to assess vulnerabilities in government Microsoft software deployments, even as other agencies remain cut off from Anthropic's tools. Some analysts have questioned whether Mythos's capabilities justify Anthropic's dramatic framing, with some studies finding that cheaper models can achieve comparable results in vulnerability discovery.

A White House official told The New York Times that talk of an executive order is "speculation," and that any announcement would come from Trump himself. Dean Ball, a former senior adviser on AI in the Trump administration, told the newspaper that officials are trying to avoid overregulation while keeping pace with the technology, calling it a “tricky balance.”

Google Preferred Source

Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.

Luke James
Luke James
Contributor

Luke James is a freelance writer and journalist.  Although his background is in legal, he has a personal interest in all things tech, especially hardware and microelectronics, and anything regulatory. 

14 Comments Comment from the forums
  • Scott_Tx
    Let's see, if the going rate for being able to sell a router in USA is 1 golden ballroom the cost for a new AI should be about... 5?
    Reply
  • chaos215bar2
    So, presumably, what they really want is to punish Anthropic for standing up to the administration and drawing a clear line where their models may and may not be used for military purposes.
    Reply
  • usertests
    Booo. Safety is for suckers.
    Reply
  • hotaru251
    this administration hasnt cared about ai regulation. just a personal dislike of what others did is why it got rid of stuff in place.
    Reply
  • timsSOFTWARE
    I wonder if the Trump administration is fully aware that the aura of "Mythos" is mostly a myth, but is trying to get back at Anthropic for non-compliance with the military deal - it seems clear to me that the "Legend of Mythos" was hoped to carry Anthropic through their IPO.

    But really, as long as AI is still based on static models that could be validated in this method, it never be able to live up to the hype. The next big step for AI would have to involve self-evolving/updating models - more akin to how human neural networks are updated every night when we sleep.
    Reply
  • American2021
    That way the government can see which ones are the best to surveil us with "automated analysis" (e.g. predictive tracking, facial recognition, backdoor methods, etc.) rubber stamped by often misused legislation like FISA which they misused and then simply sent the department heads up to lie to Congress about it afterwards. It's a lovely world.
    Reply
  • Why_Me
    chaos215bar2 said:
    So, presumably, what they really want is to punish Anthropic for standing up to the administration and drawing a clear line where their models may and may not be used for military purposes.
    Anthropic is a trash company.
    Reply
  • bit_user
    American2021 said:
    That way the government can see which ones are the best to surveil us with "automated analysis"
    No, that interpretation doesn't make sense. The models used for those purposes won't be the ones made available to the general public, which these are.

    What worries me about this is if it turns into the same thing we saw with DeepSeek models, where they refuse to answer questions about certain sensitive subjects like the 1989 Tiananmen Square crackdown.

    The administration could block models that don't agree with their worldview and could use this as a way to rewrite history. Because, too many people will apparently believe whatever AI tells them. So, if it says a certain person won a certain election or attempts to rewrite the story of slavery in the US, then that's effectively the new version of history.

    usertests said:
    Booo. Safety is for suckers.
    Yeah, I mean who wouldn't want to be subject to a terrorist attack with a chemical or biological weapon?

    That's what makes this one tricky. I think the government has a legit interest in making sure that stuff isn't too accessible. So, it's hard for me to say that they're in the wrong to want some sort of safety hurdles that models must clear.

    I guess the solution that's usually employed for sensitive security matters is for Congress to have full visibility into the regulatory process.
    Reply
  • bit_user
    timsSOFTWARE said:
    But really, as long as AI is still based on static models that could be validated in this method, it never be able to live up to the hype. The next big step for AI would have to involve self-evolving/updating models - more akin to how human neural networks are updated every night when we sleep.
    They have an ability to adapt via long context windows. That's how people are able to build "relationships" with a chatbot. Everything it remembers about their interactions is just the stuff that fits in its context window.

    Actual training is massively resource-intensive. With current technology, it wouldn't be feasible to do adaptation via individualized training.
    Reply
  • timsSOFTWARE
    bit_user said:
    They have an ability to adapt via long context windows. That's how people are able to build "relationships" with a chatbot. Everything it remembers about their interactions is just the stuff that fits in its context window.

    Actual training is massively resource-intensive. With current technology, it wouldn't be feasible to do adaptation via individualized training.
    Any context window that isn't unlimited fills up - it's just a matter of time. And as AI move beyond text and into other forms of sensory input, the equivalent of a 1M token window won't last very long at all when they are getting raw visual and audio data as sensory input, etc.

    In humans, the hippocampus is the equivalent of the context window. During each waking session, it accumulates episodic memories. Then when you sleep, the contents of the hippocampus - or at least the gist of it - gets trained into the cortex, and then the hippocampus is reset, so it can store new memories the next day. This "model fine-tuning" is the primary reason why we have to sleep.

    The problem is, we don't understand sleep very well - because we're not conscious when it happens. So it's the hardest part of the functioning of the human brain to try to replicate.

    Babies sleep the most - up to 18 hours a day - because they have the most training to do, and at that stage, it doesn't make sense to try to synchronize with the sun on a daily-basis yet. Older humans reduce the amount of sleep they need - down to as little as 6 hours - because the system is designed to reduce training/become more rigid as you get older. The reason for the latter - not as a defect, but an intentional feature - is because "catastrophic forgetting" is not just an AI model thing - learning too much, too quickly, can result in forgetting in humans as well. It's why you probably don't remember being < 3 years old. Your network was being updated drastically, and loss of memory was a side-effect. If you've been successful enough to survive into middle/older age, that's taken as a sign by the system that you learned some valuable things that should be passed down rather than forgotten.
    Reply