Washington has begun gating the release of the country's most powerful models on security grounds, and the industry that backed the president cannot agree on whether the move guards against China or hands it the lead.
The coalition that cheered President Trump's hands-off approach to artificial intelligence is now arguing with itself, in public and in real time, over a single question: whether locking down America's most capable AI models to satisfy national security keeps the country safe or simply lets China pull ahead.
The dispute broke into the open after the White House asked OpenAI to hold back a broad launch of its newest model and release it in stages, the second time in a month the federal government has stepped between an American lab and its customers.
David Sacks, who served as Trump's AI and crypto adviser before leaving the post, used X to warn that the administration was straying from its own playbook. A year ago, he noted, the president had cast AI as a race the United States would win by favoring innovation over restriction. "We deviate from that strategy at our peril," Sacks wrote.
Others see something closer to recklessness in the opposite direction. Kevin Bankston, an AI governance adviser at the Center for Democracy and Technology, told the outlet Ground Level AI that the approach was a recipe for crashing the domestic AI market.
The disagreement matters because the precedent being set now could define how the most powerful technology of the decade reaches the people who use it.
Two interventions in one month
The fight traces back to a Commerce Department order issued on June 12.
That directive required Anthropic to cut off foreign nationals from Mythos 5 and Fable 5, its two most capable systems, after researchers flagged a jailbreak vulnerability that alarmed officials. Screening users by nationality turned out to be technically impractical, and with a notice window of roughly 90 minutes, the company pulled both models offline for everyone. Mythos 5 had been built as a frontier model for cybersecurity work and shared first with a small group of vetted institutions under a program Anthropic calls Project Glasswing. Fable 5 was the more widely available version, fitted with extra safeguards.
Anthropic has since brought Mythos back for a limited set of approved U.S. organizations, after Commerce Secretary Howard Lutnick wrote in a letter that the company's cooperation with the government had produced meaningful progress. Fable may follow.
Then came OpenAI's turn. On June 26 the company released GPT-5.6 as a permissioned preview rather than a public launch, sending it first to roughly 20 partners that the government had cleared, with Amazon's Bedrock platform serving as one point of entry. Two White House offices drove the request, the Office of the National Cyber Director and the Office of Science and Technology Policy, both of which wanted to test the model's security before it spread. Lutnick phoned chief executive Sam Altman to caution him against shipping it without sign-off from other agencies. A wider release could come within weeks if the review goes smoothly.
What unsettled the industry was the mechanism Altman described to his staff: the government would be "approving access customer by customer."
OpenAI and the administration consider GPT-5.6 roughly as capable as Mythos, particularly at the kind of cybersecurity work that can find and exploit software flaws faster than any human team.
When access becomes a government decision
For years, getting a frontier model meant signing up and paying a bill. The events of June rewrote that assumption.
Enterprises that believed they were buying a service discovered their access could be interrupted by a federal order, a category of risk most procurement teams had never modeled. OpenAI made clear it does not want the arrangement to harden into routine, warning that case-by-case clearance keeps the strongest tools away from the developers and defenders who need them most.
There is, as yet, no formal rulebook behind any of this. Earlier in the month Trump signed an executive order on advanced AI that asks companies to submit frontier models for a government cybersecurity review up to 30 days ahead of release, though participation is voluntary, and Google, Microsoft, xAI and others had already agreed to let officials test their models beforehand. The Anthropic shutdown came through Commerce, while the OpenAI request came from the White House, a split that has left companies unsure which part of the government is actually setting the terms.
Lawmakers have noticed the vacuum. A bipartisan group of House members wrote to Lutnick on June 18 questioning the legal basis for the Anthropic controls, in particular whether export-control authority extends to running a model through an API, and set a deadline for a written justification.
Brad Carson, who leads the group Public First, summed up the worry among many executives. Recalling a dangerous product is fair game for the government, he argued, yet the current setup amounts to an "ad hoc, personalized, opaque, possibly lawless approach."
The China problem the controls are meant to solve
The security camp is not arguing in a vacuum, and the evidence on the other side of the ledger is mounting.
The Wall Street Journal reported that Chinese systems have matched Anthropic's Mythos in some cybersecurity scenarios, a shift with the potential to reset the global technology race. The immediate cause for concern is GLM-5.2, a model from the Chinese developer Zhipu AI. Two independent evaluations, by the security firms Graphistry and Semgrep, found it performing on par with leading American models at investigating threats and discovering vulnerabilities. Graphistry went further, suggesting the model might be an illegal distillation of GPT-5.5 and Opus 4.8, which would help explain how fast the gap has closed.
Here is the part that makes a government speed limit so awkward. GLM-5.2 is open-weight, meaning anyone can download it and run it on their own hardware with the safety controls removed and no provider watching. A closed model can be switched off, as Anthropic's was. An open one cannot.
Developers are voting with their workloads. Chinese open-weight models have gone from a rounding error on the routing platform OpenRouter in late 2024 to a commanding share of its traffic this year, propelled by aggressive pricing as companies hunt for ways to trim their AI bills. By one tally of usage in late June, Chinese models accounted for nearly half of the platform's tokens while the American share had slipped to about a fifth.
The competitive anxiety carries a darker edge. Anthropic has alleged that several Chinese labs, among them DeepSeek and Moonshot AI, ran industrial-scale campaigns to distill its Claude models, and two House committees opened a joint investigation in late April into the national security risks posed by Chinese AI.
The case for clear rules
For all the noise, a striking number of the people caught in the middle are asking for the same thing: clarity.
Anthropic itself has pushed for stronger safeguards as models grow more dangerous, putting it closer to the administration's instincts than to Sacks's. Mark Pincus, the Zynga founder who has invested in both OpenAI and Anthropic, said he backs clear regulation while noting how hard it is to build against a moving target. Dan Shipper, who runs the AI subscription service Every, called the government's involvement important and said the task now is to strike the right balance between safety and access.
The catch is that rules require yardsticks, and yardsticks can be gamed. Siméon Campos, an AI startup founder, cautioned that labs could engineer their systems to clear whatever benchmarks regulators adopt.
Investors, meanwhile, are repricing the risk. The venture capitalist Paul Kedrosky called the situation bearish for the labs, describing a market that suddenly has a hall monitor watering down the punch bowl, the sort of uncertainty that drags valuations lower when a company's best products can be held at the government's discretion.
Box's Aaron Levie put the broader stakes plainly, calling the moment "de facto AI regulation." For most of the past two years, he argued, the constant leapfrogging between rival labs is what pushed the technology forward so quickly.
The question the House put to Lutnick still hangs over the industry. The deadline for a written explanation of the Anthropic controls passed with no public response from Commerce, leaving the American labs and their customers to guess where the line sits and who gets to draw it.
Comments
Join the discussion and share your perspective.