The Agentic Post
Breaking
Black October 2026: Every Major AI Security Incident, Explained  ·  ARTEX: The Open-Source AI Hacking Tool Behind October’s Biggest Attacks  ·  Chinese Hackers Used AI Agents to Hit 2,250 Microsoft 365 Accounts  ·  PHP2Shell Is Now Being Exploited, and Attacks Began Before the Patch  ·  OpenAI-Backed Thrive Buys ADT for $15.3 Billion to Rebuild It With AI  ·  The AI Data Center Backlash Hits Minnesota and Australia in One Week  ·  
Home/AI Models/DeepSeek
US Orders DeepSeek Off Federal Systems, and China Threatens to Respond

US Orders DeepSeek Off Federal Systems, and China Threatens to Respond

DeepSeek

DHS ordered civilian federal agencies to remove DeepSeek and any third-party service built on it, citing data risks. China's Ministry of Commerce called the ban discriminatory and warned of countermeasures.

The US Department of Homeland Security has ordered all civilian federal agencies to remove DeepSeek from their computer systems and stop buying it, including any outside product or service built on top of DeepSeek. The order was filed publicly on October 19, 2026 and published in the Federal Register on October 20. China’s Ministry of Commerce responded with what NBC News described as "strong dissatisfaction and firm opposition," calling the ban discriminatory and warning Beijing would take "all necessary measures."

What does the order actually require?

Known as Exclusion and Removal Order 2026-01, it does two things. It removes DeepSeek and its affiliates’ products from civilian agency systems, and it bars agencies from buying them, including through approved product lists and online marketplaces. Crucially, it also covers third-party products and services "developed using" DeepSeek. The order cites the Federal Acquisition Supply Chain Security Act of 2018 and follows a recommendation from the Federal Acquisition Security Council. The order itself is dated July 24, 2026, so it was signed months before it was made public.

Why does DHS say DeepSeek is a risk?

DHS says DeepSeek’s models present "multiple threat vectors." It points to Chinese laws that let the government access private company infrastructure, account sign-ups that collect personal information and may fingerprint users, and the fact that using any chatbot means handing it your information. DHS warned use "may result in the transfer of sensitive data to the PRC" and "could produce actionable intelligence on U.S. persons and/or government agencies," and added that the models themselves may contain malware or other vulnerabilities.

Is this the first restriction on DeepSeek?

No, but it is the broadest. The Commerce Department and the US Navy had already banned DeepSeek on government devices, and New York, Tennessee and Virginia took similar steps. Italy, Australia and Taiwan moved early outside the US. What is new is the reach to third-party services built on the model, and the formal procurement ban across every civilian agency.

What about the fact that DeepSeek is open source?

This is the part that makes the order hard to apply. DeepSeek releases open weights, like V4.1 Flash under the MIT licence, and anyone can run them on their own servers with no connection to China. The data-transfer risk mainly applies to DeepSeek’s own hosted service on Chinese servers. A model running entirely on a US government server does not send anything to Beijing. But "developed using DeepSeek" is broad enough that vendors who fine-tuned or built on its open weights may now need to prove they did not, or remove those products from federal sale.

What happens next?

China has not said what its countermeasures would be. The order lands in a wider push against Chinese technology, with the FCC voting on October 29 on barring Chinese labs from testing electronics for the US market. It also sits awkwardly beside the industry trend toward cheaper open models, documented in AT&T moving traffic to open models, since several of the strongest open models now come from Chinese labs. For businesses selling to the US government, the practical step is an audit: know which models sit underneath every AI tool you supply.

See Decrypt’s report on the order.

Up Next
LiteLLM 1.107.0 on PyPI Steals AI Keys and Opens a Backdoor

LiteLLM 1.107.0 on PyPI Steals AI Keys and Opens a Backdoor

AI Safety

LiteLLM 1.107.0 on PyPI ships a credential stealer targeting cloud, SSH and AI tool keys plus an hourly backdoor. It runs whenever Python starts. Versions 1.106.2 and the release candidates are clean.

LiteLLM version 1.107.0, published to PyPI, contains a hidden credential stealer and a backdoor, security firm SafeDep reported on October 21, 2026. LiteLLM is a widely used Python library that lets apps talk to many AI model providers through one interface, so the machines it runs on often hold API keys for OpenAI, Anthropic and others. The malicious files run every time Python starts on the infected machine, even if your code never imports LiteLLM. Versions 1.106.2, 1.107.0rc1 and 1.107.0rc2 are clean.

How did malware get into an official release?

Through the build, not the source code. The code on LiteLLM’s public GitHub repository is clean. But the package uploaded to PyPI had one extra line in its build settings that copied a hidden folder of malicious files into every installer, on Linux, macOS and Windows. SafeDep concluded that someone tampered with the build outside the public repository, either through a compromised release pipeline or stolen PyPI credentials.

That is what makes this kind of attack hard to spot. Reviewing the code on GitHub would show nothing wrong. Only inspecting the actual package you downloaded reveals it.

What does the malware steal?

Almost everything worth stealing from a developer machine or server. The stealer collects SSH keys, AWS, Azure and Google Cloud credentials, Kubernetes and Docker configs, saved browser passwords, shell history, CI/CD secrets and private keys. It specifically targets AI developer tools, including credentials and configs for Claude Code, Claude Desktop, Cursor, Copilot, Codex, Gemini CLI, Windsurf and MCP configuration files. If it finds a LiteLLM proxy database, it dumps the stored AI provider API keys.

It encrypts the haul and sends it to a lookalike domain, www.litellm[.]dev. A second file installs a background process that, every hour, checks a GitHub file for a server address, steals credentials again and downloads and runs whatever code that server sends. In plain terms: once installed, the attacker can run anything they like on the machine, indefinitely.

What should you do if you use LiteLLM?

  • Check your version: run pip show litellm. If it says 1.107.0, treat the machine as compromised.
  • Pin to a clean version such as 1.106.2 and reinstall.
  • Look for the implant files cook.pth, cook.py and sitecustomize.py in your Python site-packages folder, and the marker files /tmp/.pc_once_72129118c59f5636c0bb and /tmp/remote_code_loader.lock.
  • Kill any detached Python background process you did not start.
  • Rotate everything: cloud keys, SSH keys, AI provider API keys, GitHub tokens, and any passwords saved in browsers on that machine. Uninstalling the package does not undo theft that already happened.
  • Check CI pipelines, since the stealer runs inside build jobs too.

Why is this part of a bigger pattern?

AI tooling has become the most valuable target in the software supply chain, because the machines running it hold the most powerful keys. The same day, hosting platform Railway disclosed a breach it linked to a wave of attacks on platforms that host AI inference. It follows slopsquatting fake packages, GitSpawn and AgentCorruption. The lesson is the same each time: pin your dependency versions, verify what you install, and keep AI credentials scoped so one stolen key cannot unlock everything.

Read SafeDep’s full analysis, including file hashes.

Part of our coverage of Black October 2026, every major AI security incident this month.