DHS Bans DeepSeek from Federal Systems

DHS Secretary Markwayne Mullin signed FASCSA Exclusion Order 2026-01 on July 24, 2026, but the directive stayed out of public view until it surfaced in the Federal Register on October 19-20, 2026, a roughly three-month gap between signing and disclosure ^[1]. That delay is notable less for secrecy than for precedent: the order isn't a novel policy tool but an application of authority Congress created in 2018, when it passed the Federal Acquisition Supply Chain Security Act and stood up the Federal Acquisition Security Council as a standing mechanism to exclude risky vendors from the federal supply chain without needing new legislation each time ^[2]. That same authority was previously invoked against Huawei, ZTE, Kaspersky Lab, and Chinese surveillance-camera makers between 2019 and 2021, which places DeepSeek in a well-worn lineage of China-linked exclusions rather than marking a genuinely new approach to AI policy ^[2].

The three-month silence suggests agencies were already quietly working through removal and exception processes before the public, or DeepSeek itself, had any visibility into the order's existence. Framed that way, the 'ban' reported this week is really the disclosure of a decision that has already been shaping federal procurement behavior since midsummer.

The exclusion's reach is its most consequential design choice. Rather than banning only DeepSeek-branded software, it extends to any 'third-party products or services developed using DeepSeek implementations,' closing the obvious workaround of wrapping or rebranding the underlying model before selling it to a federal buyer ^[3]. Gabrielle Hempel, a security operations strategist at Exabeam, called that breadth the order's most striking feature: 'What I find particularly interesting about the DeepSeek order is the breadth of its language' ^[4]. She also pushed back on the premise that national origin alone should drive the risk calculus: 'I understand the national security concerns, particularly around data sovereignty and foreign intelligence collection, but country of origin isn't necessarily a proxy for security risk' ^[4].

Agencies that believe they genuinely need DeepSeek-derived technology aren't simply out of luck - they can file an exception request, but only 60 days in advance and only with a 'compelling justification,' such as proving no substitute tool can fulfill a mission-critical function ^[1]. DHS grounded the underlying risk case in two distinct arguments: a data-sovereignty problem, since DeepSeek stores user data on China-based servers where national security law gives Beijing 'wide latitude to request and obtain data' ^[1], and a separate, more mundane technical case - multiple critical vulnerabilities identified in DeepSeek's iOS app, plus an account-creation flow that DHS says can collect personally identifiable information and digitally fingerprint users ^[5]. Hempel's objection points to the ambiguity this breadth creates in practice: agencies and contractors now have to audit model lineage across their software supply chains, not just check whether 'DeepSeek' appears in a product name.

The DHS order doesn't exist in a vacuum - it sits alongside a tougher legislative push that actually predates it. Sen. Josh Hawley (R-MO) introduced S.321, the 'Decoupling America's Artificial Intelligence Capabilities from China Act,' on January 29, 2025, just days after DeepSeek's R1 model launch intensified US-China AI competition concerns ^[6]. Unlike DHS's agency-level procurement order, Hawley's bill would reach far beyond federal systems - reporting on the bill describes it as criminalizing broader US-China AI transactions, including individual Americans' use of models like DeepSeek R1 ^[7]. More than a year and a half later, the bill remains stalled in the Senate Judiciary Committee and hasn't become law ^[6], which makes the DHS executive order the first piece of this agenda to actually take effect rather than merely being proposed.

The contrast is instructive: DHS acted through existing administrative authority it already had, while Congress's more sweeping ambitions, potentially touching private citizens rather than just government systems, remain stuck in committee. That gap between what the executive branch can already do unilaterally and what would require new legislation is likely to keep showing up as US-China AI policy continues to evolve.

Beijing didn't let the order pass quietly. China's Ministry of Commerce said the ban disregarded an understanding reached between Xi Jinping and Trump and amounted to using 'national security' as a pretext to suppress Chinese enterprises, framing the move as politicized and disproportionate ^[8]. China's Global Times carried similar language from government channels warning the country will 'take all necessary measures' to protect its firms' interests, raising the prospect of retaliatory measures against US companies operating in China ^[9].

On the commercial side, the order's effects are likely to radiate well past the federal government it legally binds. Enterprise security teams now have a concrete federal precedent to cite when flagging DeepSeek-derived technology in their own vendor risk reviews, even though nothing requires them to follow DHS's lead ^[3]. Because forced migration favors vendors that already hold completed compliance paperwork, AI providers like OpenAI, Anthropic, and Google stand to gain by default, not necessarily because federal buyers re-evaluated their models against DeepSeek's, but because agencies tend to default to whichever vendor already has finished authorization paperwork rather than re-benchmarking alternatives from scratch ^[3].