CISA Accuses DeepSeek of Copying US AI Models as DHS Readies $626M Contract
On Sept. 8, 2026, CISA published advisory AA26-251A on China-based AI companies. CISA That advisory is the lead federal cybersecurity warning in this set of actions.
The advisory states DeepSeek ran an organized distillation campaign against frontier AI models built by U.S. companies since at least late 2024. Distillation means using the outputs of a powerful, costly model to train a cheaper copy. Frontier models are the most capable systems currently in use. The start date is no later than late 2024, and CISA frames it as a campaign, not a single incident.
An earlier proposal dealt with DeepSeek on federal networks. House lawmakers proposed banning DeepSeek on U.S. government devices, according to Feb. 6, 2025 reporting. Supporters cited the Chinese government's ability to use the app for surveillance and misinformation as the reason to keep it off federal networks. PBS NewsHour The proposal applied to government devices and federal networks, not to private-sector use.
On procurement, DHS expects to award the potential $626 million Network, Cloud and Cybersecurity Services 2.0 contract by late November, according to Oct. 2, 2026 reporting. Federal News Network The contract covers network, cloud and cybersecurity services. The $626 million figure sets the near-term pipeline for vendors in those areas.
CISA has also issued three directives on basic agency hygiene. It issued Emergency Directive ED 25-03 in response to an ongoing exploitation campaign by a sophisticated threat actor. On Feb. 5, 2026, it required removal of unsupported hardware and software to improve security of edge devices, the internet-facing gear like routers and firewalls. CISA On June 10, 2026, it issued Binding Operational Directive BOD 26-04, Prioritizing Security Updates Based on Risk, to consolidate rules for fixing vulnerabilities. CISA
The three directives use different tools. ED 25-03 answers active exploitation. The Feb. 5 directive mandates removal of unsupported products. BOD 26-04 standardizes how agencies rank security updates by risk. Together they tighten patching, asset tracking and replacement schedules across federal civilian agencies.
For savers and investors holding AI stocks and bonds, the advisory widens what counts as stealable. Model weights and chatbot outputs become assets that can be systematically collected. Distillation at scale, if confirmed as described, lowers the return on expensive training runs and makes it harder to trace where a downstream model came from. That shifts attention to access controls, rate limits, terms-of-use enforcement and tracking of bulk use, not just firewalls.
The broader context here is sustained federal demand for cleanup work. A $626 million services contract, plus orders to purge outdated edge gear and reprioritize patches by risk, points to ongoing compliance work for agencies and contractors. Vendors focused on continuous monitoring, vulnerability management and secure cloud moves may see task orders, while agencies face costs to inventory and replace end-of-life equipment. The older, narrower ban proposal shows lawmakers remain willing to restrict specific apps on federal devices on counterintelligence grounds.
