Trump killed his own AI security order. The agency that was supposed to act on it cant access the models.
Trump killed his own AI security order. The agency that was supposed to act on it cant access the models.
The signing ceremony was scheduled. Then it wasnt.
President Trump postponed his executive order on AI oversight Thursday morning, pulling it hours before the ceremony for what he called a simple reason: I didnt like certain aspects of it. Speaking in the Oval Office, he was more specific: I dont want to do anything thats going to get in the way of that lead. The order, which draft details published by POLITICO showed included a 90-day voluntary pre-release review window for frontier AI models and provisions for government access to systems like Anthropics Mythos, was dead on arrival.
The post-mortem will focus on White House infighting, which Axios reported centered on disagreements over whether the order was too restrictive or not restrictive enough. That is the news. It is not the story.
The story is the gap between what the government says it wants to do about frontier AI security and what it is actually capable of doing. That gap just got wider.
Here is what the draft order would have required, according to POLITICO: developers of covered frontier AI systems would have had to submit their models to a voluntary review process involving the NSA, Treasury, and CISA as many as 90 days before public release. The NSA would have made final determinations on which models triggered the requirements. CISA would have been in the room.
CISA is not in the room. CISA does not have access to Mythos.
That is not a processing delay. Axios reported in April that the Cybersecurity and Infrastructure Security Agency, the civilian agency designed specifically to defend critical infrastructure from cyber threats, has no access to Anthropics Mythos Preview model. The same model whose cyber capabilities Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell found alarming enough to convene an urgent meeting with the CEOs of JPMorgan, Goldman Sachs, Morgan Stanley, Bank of America, Citigroup, and Wells Fargo in April.
CISA is not excluded because it chose not to engage. Anthropic told CNBC it has been in ongoing discussions with CISA about Mythos cyber capabilities. Those discussions have not produced access.
The agencies positioned to oversee frontier AI are the ones with seats at the table. CISA, the civilian clearinghouse for critical infrastructure cyber risk, is not at the table. And the workforce that would have done the technical evaluation has been substantially reduced: CISA has lost more than a third of its staff under the Trump administration.
The NSA is in the room. The NSA, according to the draft EO, would have made final determinations on which AI systems constituted covered frontier models. The agency is not starting from zero on this question. NSA has been testing Mythos to find cybersecurity vulnerabilities in Microsoft products, with officials expressing impressed by its speed and efficiency, according to people familiar with the matter. NSA is acting in its offensive security mission, not as a civilian evaluator, and it is doing so through a private arrangement with Anthropic that bypassed the standard government access process entirely.
Goldman Sachs CEO David Solomon confirmed on the banks earnings call that Goldman has the model. Morgan Stanley CEO Ted Pick said the same. Deutsche Bank CEO Christian Sewing told Reuters that every bank is trying to gain access. The British governments AI Security Institute went further than any U.S. agency, publishing an assessment that Mythos is substantially more capable at cyber offence than any model we have previously assessed.
CISA has not published an assessment. CISA cannot read the model.
Project Glasswing is, in effect, a private-sector AI oversight regime. Anthropic decided which partners got access, which banks were briefed, which governments were engaged. The labs are writing the rules that the government is still trying to figure out how to enforce.
Trump said he postponed the order because it could be a blocker. He may sign a revised version. The 90-day review was voluntary. Nothing in the draft would have forced Anthropic, OpenAI, or any other lab to share their most powerful models with CISA if those labs chose not to.
What the order would not fix, revised or not, is the structural problem: the agency with the statutory mandate to defend critical infrastructure from AI-enabled cyber threats does not have the people or the access to do the job. The agency with the workforce and the intelligence relationships is the NSA. The civilian cyber defense mission is now running on a skeleton crew that cannot read the threat models it is supposed to be watching.
The next time a model like Mythos drops and the Treasury Department needs to warn banks about systemic risk, the government will have to do it without the technical picture that CISA should have been building for the past two years.