Opinion: Washington just banned DeepSeek for the third time in ten months. That's the real failure.

The Department of Homeland Security's order barring federal civilian agencies from buying or running DeepSeek products, formally published Tuesday in the Federal Register, reads like decisive action: immediate removal of a Chinese artificial-intelligence company's software from government systems, a 60-day window for agencies to seek exceptions, and the Cybersecurity and Infrastructure Security Agency standing ready to review them. What it is not is new. It is the third time in ten months that Washington has had to write a separate rule to keep the same company's technology off a different slice of the federal government — and the gaps between those three rules say more about the state of federal AI governance than any one of them does on its own.

Three bans, three authorities, ten months apart

Start with the timeline. In December 2025, Congress folded a provision into that year's defense policy bill — now Public Law 119-60, the National Defense Authorization Act for Fiscal Year 2026 — that barred the Pentagon and its contractors from using DeepSeek's models, a restriction that took effect Jan. 17, 2026 under the law's Section 1532. A separate provision in the same mammoth bill, Section 6604, imposed a parallel ban specifically on intelligence-community systems. Both were narrow, deliberate, and passed through the ordinary legislative process, with debate, committee review and a presidential signature.

Nothing comparable happened for the rest of the federal government — the Department of Education, the IRS, the Social Security Administration, the National Park Service — until this week, when reporting on the new order made clear civilian agencies had been operating for ten months under no DeepSeek-specific restriction at all, and even now the fix didn't come through Congress. DHS reached instead for the Federal Acquisition Supply Chain Security Act of 2018, a law written to let the government exclude specific vendors from federal procurement on supply-chain-security grounds, acting on a recommendation from the Federal Acquisition Security Council. It is a legally sound mechanism. It is also a tool built for a different, narrower problem — vetting individual contractors and suppliers — now being stretched to cover an entire emerging category of commercial software that didn't exist when the statute was written.

There was, in between, an attempt to do this the straightforward way. Reps. Josh Gottheimer and Darin LaHood introduced the No DeepSeek on Government Devices Act in February 2025, which would have had the Office of Management and Budget set a single government-wide standard for removing DeepSeek from federal information technology — one rule, one authority, covering every civilian agency at once. It never got out of committee. Nearly two years later, DHS has had to approximate what that bill would have done anyway, using a different law, covering a narrower definition of "federal," and arriving nearly ten months after the Pentagon's version of the same ban.

The pattern, not the policy, is the problem

None of this is an argument that DeepSeek should be running on federal systems. DHS's stated rationale — detailed in the order itself — cites a 2025 security researcher's findings that DeepSeek's mobile app shared user data with ByteDance, concerns about Chinese government access to the company's infrastructure, and the possibility that government data processed by DeepSeek's models could be exposed. That reasoning is specific, documented and not seriously in dispute. A reasonable government bans software like that. The question worth asking is why banning it took three separate legal maneuvers spread across ten months, rather than one.

The honest answer is that the federal government does not have a standing mechanism for evaluating foreign AI systems before they spread through agency IT and then deciding, government-wide, what to do about them. It has a patchwork: a defense authorization bill that happens to pass every year and can carry AI provisions when lawmakers remember to write them in; an intelligence community with its own separate statutory lane; and, for everyone else, a 2018 supply-chain-security law originally aimed at hardware vendors and IT contractors, pressed into service because nothing more fitting exists. Each of those authorities has its own definition of "federal," its own timeline, its own exception process. DHS's order this week applies, by its own terms, only to civilian executive agencies — not to state governments (a handful, including Texas, have imposed their own bans since January 2025), not to federal contractors operating outside DOD's reach, and explicitly not to private businesses or individual Americans, who make up the overwhelming majority of DeepSeek's actual user base in this country.

That fragmentation matters because DeepSeek will not be the last foreign AI system to raise this exact set of concerns. The next one — built by a different company, in a different country, with a different app and a different vulnerability — will arrive to find the same absence of a standing framework. Someone will have to decide, again, which of several ill-fitting legal authorities to invoke, for which slice of government, on what timeline. If the DeepSeek sequence is any guide, that process will again take the better part of a year to reach anyone outside the Pentagon.

A fair counterargument is that speed matters less than legal footing, and that DHS simply used the fastest authority actually available to it rather than waiting on a Congress that let the cleaner bill die in committee. That is true, and it is also the point: the fastest available authority was a 2018 supply-chain law never written with large language models in mind, applied through a council whose public track record, before this week, consisted mostly of hardware and telecom-equipment exclusions. The order's own compliance mechanics show the strain. Agencies get 60 days to request an exception, and CISA — already responsible for reviewing supply-chain risk across the entire civilian government, defending federal networks, and running the nation's principal cyber-incident response function — now also adjudicates those requests one by one, agency by agency, for a single banned company. Multiply that by however many foreign AI products eventually warrant the same scrutiny, and the review queue becomes the bottleneck the current system was never designed to avoid.

What a workable answer would look like

A government-wide framework does not require inventing new law from scratch. The Federal Acquisition Security Council already exists, already has a statutory mandate to recommend exclusions, and already proved this week that it can act on short notice. What it lacks is a standing, published methodology — public criteria for what makes a foreign AI product a supply-chain risk, a routine review cadence rather than a one-off response to a news cycle, and explicit coverage that extends uniformly across civilian agencies, the intelligence community and the Pentagon under a single determination rather than three. Congress could supply that coverage directly, as the stalled 2025 bill proposed, or it could simply instruct FASC to apply one standard everywhere rather than leaving DOD and the intelligence community to legislate their own carve-outs while everyone else waits on a different agency entirely.

The alternative is what just happened: a real and documented security risk that took root across the government for most of a year before civilian agencies got the same protection the Pentagon had already received in January. DHS's order is not wrong. It is simply proof that the only standing policy for foreign AI risk in the federal government is to improvise one, company by company, after the fact — and that each time, it will take months longer to protect most of the government than it took to protect the parts Congress happened to remember.