🏛️ DHS (DeepSeek ban), ⚡ DeepSeek (DeepInfer) and 🛡️ Microsoft (Guardian Agents) — trust is not free
DHS orders civilian federal agencies to remove DeepSeek and any third-party tools built on its models. The same day DeepSeek open-sources DeepInfer, the inference engine behind its production services. Microsoft declares that no frontier model can be trusted and launches Guardian Agents to watch them like rebellious employees.
Washington cuts the Chinese stack while DeepSeek gives away the infrastructure that makes its service cheap, and Microsoft admits agents are already a real internal threat.
🏛️ DHS bans DeepSeek from civilian federal systems
The Department of Homeland Security made public on October 20 the order that excludes products from Hangzhou DeepSeek Artificial Intelligence from civilian federal purchases and requires the removal of any DeepSeek product or third-party tool developed using its implementations. The measure is based on the Federal Acquisition Supply Chain Security Act of 2018 and covers all civilian executive agencies. It does not cover the Department of Defense or the intelligence community.
DHS cites “multiple threat vectors”: possible transfer of sensitive data to the People’s Republic of China, generation of actionable intelligence on people and U.S. agencies, and the risk that the models themselves contain malware or vulnerabilities. The order, signed in July and made public now, closes the loophole of using DeepSeek through an intermediate vendor.
⚡ DeepSeek opens DeepInfer under Apache 2.0
DeepSeek published DeepInfer on October 20, the inference framework that powers its production services for the V4.1-Flash model, under the Apache 2.0 license. The code targets Nvidia Blackwell GPUs and Huawei Ascend NPUs. It includes optimized kernels, distributed parallelism, speculative decoding and structured outputs. It leaves out the KV cache system and load balancing, which remain tied to its internal infrastructure.
The stated goal is to lower compute costs and make AI more affordable. The timing with the U.S. ban is not accidental: the same day Washington cuts access, DeepSeek gives away the engine that makes its service hard to compete with on price.
🛡️ Microsoft: frontier models cannot be trusted
Hayete Gallot, Microsoft’s head of security, wrote on October 19 that “frontier models, regardless of their origin, cannot be trusted.” Not because they are malicious or easy to hack, but because they are unpredictable: they are trained to achieve outcomes, not to follow an organization’s rules. Microsoft presents Guardian Agents inside Project Perception to monitor, restrict and contain other agents.
Agents have broad access to data and systems and can collude with each other without anyone compromising them. Microsoft treats them as “agentic insiders” and offers controls to require human approval or shut them down. The admission is clear: the risk is no longer only outside, it is inside the company.
The connecting thread
The U.S. government cuts access to a Chinese lab while that same lab opens the engine that makes its service economically hard to match. At the same time Microsoft, which sells agents to enterprises, admits it cannot trust any frontier model and needs another agent to watch them. Trust, sovereignty and cost are being negotiated at once.