Emil Michael, the Pentagon's chief technology officer, says Washington can use export controls on AI companies that do not cooperate with the government, and that it has already done so. He made the case in a video interview with Amir Efrati and Leo Schwartz of The Information, published Monday, October 19, 2026, and summarized by The Next Web.
Some of what he said is a policy position, which is easy to report accurately. Some is a factual claim about a past event involving Anthropic, which is not independently confirmed in anything we could find. This post keeps the two apart.
Quick facts
| Question | Answer |
|---|---|
| Who said it? | Emil Michael, Pentagon CTO and vice chair of the White House Super Intelligence Force |
| Where? | A video interview with The Information, reported October 19, 2026 |
| Core statement | Export controls are an appropriate tool for AI companies that do not cooperate |
| Precedent he cited | Past limits on exporting encryption; limits on foreign nationals using a technology |
| Biggest unverified claim | Anthropic refused to pull back a jailbroken model and was under export controls within two hours |
| Policy outputs he expects | An incident response framework, plans to harden critical infrastructure, a review of existing laws |
| Anthropic's status | Labelled a supply chain risk; asked the full appeals court to rehear its case |
What he actually said
According to The Next Web's account of the interview, a White House official had named export controls as one possible stick for the labs. Michael was asked whether they were an appropriate way to keep companies in line and said they were. He pointed to the encryption export limits of the past and to limits on foreign nationals using certain technology. He added that the executive branch has other tools, including enforcement by the Justice Department.
He said the administration would act with urgency if a real AI problem emerged. He criticized AI companies that call for regulation, saying the reasoning does not hold up. He also said existing laws are "more than sufficient," based on his own reading and on FTC chair Andrew Ferguson's views.
The claim that is not verified
Michael said the government had already shown it would act: according to him, Anthropic refused to pull back a model after the government found out it was jailbroken, and within two hours the company was under export controls.
We could not corroborate this. The search results and the coverage we read repeat it as his account. They do not link a public order, a notice or a company statement. In the coverage we read, nothing describes which model, which controls, or what "pulled back" meant. Treat it as a claim by an official who is party to an ongoing dispute with Anthropic. That does not make it false. It means readers should not cite it as established.
Our earlier coverage of the dispute shows why the context matters. The Pentagon labelled Anthropic a supply chain risk after contract talks failed. A district-level ruling found the blacklist illegal, as we covered in the judge's August ruling, and then an appeals decision went the other way, as in the D.C. Circuit overturning that ruling. Anthropic has described the dispute as centering on two exceptions, mass surveillance of Americans and fully autonomous lethal weapons; Michael has said Anthropic wanted approval rights over particular uses.
Where this fits: the Super Intelligence Force
Michael spoke the day after the first full meeting of the White House AI task force. We covered that meeting in Super Intelligence Force first meeting: PCAST report and lab early-warning protocols. As vice chair, Michael is responsible for the government's relationship with the labs and for how they report incidents, which links to the White House incident-disclosure mandate.
Other pieces of the same policy thread:
- Liability, not new rules. The task force has leaned toward asking who pays when agents go wrong, as in Trump's AI task force on liability.
- Bessent's "slow down." Treasury's Scott Bessent urged a slower pace in the earlier task force coverage.
- Chip and remote-access controls. Export controls on hardware and remote access were already in play, as in August's chip export-control and remote-access reporting.
The new element is the suggestion that export controls could be pointed at a US lab's own model, not only at sales to China. Export controls on model weights have a legal history (encryption is the analogy Michael chose), but applying them to a domestic company's deployment would be a major step. We found no draft rule or notice describing one.
Why the encryption analogy cuts both ways
Michael's chosen precedent is worth a closer look. In the 1990s the US treated strong cryptography as a munition, so shipping it abroad required a license. The rules were challenged in court on free-speech grounds and largely loosened around 2000, after industry argued that foreign buyers would simply use products from other countries. Supporters of the analogy say it proves export law can reach software and ideas. Critics say it proves the opposite: controls on widely reproducible software tend to push users and talent elsewhere and erode the controlling country's market position.
AI weights are a closer fit to the crypto case than to chips. A chip is a physical object that can be tracked at a border. A model is a file that can be copied, hosted abroad, or reproduced by rivals, which is why the debate over controls on open weights has been so heated, as in our piece on crypto-era export controls and open weights. Using the same tool as a disciplinary lever against a domestic lab, rather than as a shield against foreign buyers, would be new. That is why the details of any real order matter far more than the rhetoric.
Other statements worth noting
Michael also said, per The Next Web:
- Anthropic and the Department. Anthropic is out of the Department of War, as the department said six months earlier it would be. He expects the government to win in the appeals court. He still talks to the company and says he will treat it fairly in task force work.
- The sticking point. He said there is no discussion of Anthropic's return unless it changes its policy that a model cannot override a commander's chain of command. He noted that Musk's company and Google accepted the same all-lawful-uses terms.
- Adoption. About 1.7 million of the department's 3.5 million employees use AI, up from 80,000 a year ago; the Pentagon's AI platform had 1.5 million users in June. These are his numbers.
- Open models. He favors American open-source models, naming Nvidia's Nemotron, Reflection AI and Thinking Machines, though none is at the frontier yet, and he worries a little about Chinese open models becoming dominant in the US. See Reflection's open-weight release and, the same day, DHS ordering DeepSeek out of federal civilian systems.
What is claimed vs what is verified
| Claim | Source | Status |
|---|---|---|
| Export controls are an appropriate tool for uncooperative labs | Michael, via The Information | Confirmed as his stated view |
| Executive branch also has DOJ enforcement and other tools | Michael | Confirmed as his stated view |
| Anthropic refused to pull back a jailbroken model and was under export controls in two hours | Michael | Unverified; no public document found |
| Existing laws are "more than sufficient" | Michael, citing FTC chair Ferguson | His reading; not an FTC statement we found |
| 1.7 million of 3.5 million employees use AI | Michael | Self-reported figure |
| Pentagon labelled Anthropic a supply chain risk | Public record and prior coverage | Confirmed; litigation ongoing |
How to check this yourself
- Read the original interview at The Information (paywalled) and compare it with the Next Web summary.
- Search the Federal Register and the Commerce Department's Bureau of Industry and Security notices for any rule naming a US AI model or lab. Our search found none.
- Look for a statement from Anthropic. Its public position on the supply chain designation is the best context for evaluating Michael's account.
- Watch for the task force outputs Michael listed: the incident response framework and the review of existing laws.
What this means for people who build with AI
- Model access risk is rising on paper. If export controls can be used against domestic labs, enterprises that depend on a single provider should have a fallback model and a contract clause for sudden unavailability.
- Incident reporting is coming. Whatever the stick, the task force is building a reporting framework. Labs and large deployers should expect to log and disclose serious AI incidents.
- Government work is shaped by "all lawful uses" terms. Vendors selling to defense agencies are being asked to accept those terms; others may wish to review how their usage policies interact.
- Security of agents matters in either case. For teams deploying agents, tools like AgentBeam, the agent security platform from the explainx.ai team, focus on stopping agents before they take dangerous actions.
This post reflects reporting available on October 19, 2026. The two-hour export-control claim is unverified and may be confirmed or contradicted by later reporting.