DeepSeek banned from U.S. federal civilian agencies under new DHS order

Original Article ↗ Hacker News Discussion ↗

U.S. Homeland Security has ordered federal civilian agencies to remove and stop procuring AI products based on Chinese firm DeepSeek’s models, citing risks of Chinese state access, data exfiltration, and potential backdoors even when models are self‑hosted. Commenters are split between viewing this as a reasonable national security precaution given opaque LLM behavior and China–U.S. tensions, and seeing it as protectionism and regulatory capture aimed at shielding U.S. AI vendors from a cheaper, highly capable open‑weight competitor. The move raises broader questions about how much any government should trust foreign AI models, whether “open weights” meaningfully improve security, and how such bans will shape the future availability and pricing of open models.

Scope and Meaning of the DHS Ban

  • Order targets U.S. federal civilian agencies, requiring them to:
    • Exclude DeepSeek and affiliated entities from procurements.
    • Remove DeepSeek products and “third-party products and services developed using DeepSeek implementations” from IT systems.
  • Major ambiguity: what counts as a “DeepSeek implementation.”
    • One reading: any instance of DeepSeek models, including self‑hosted open weights.
    • Another reading: only products/services directly using DeepSeek‑provided infrastructure or commercial offerings.
  • Several commenters note the text and footnotes are vague enough to scare contractors away from DeepSeek entirely, and potentially allow selective enforcement.

Security and Backdoor Concerns

  • Stated DHS rationale: PRC law allows state access to company infrastructure; DeepSeek collects PII, can fingerprint users, and servers in China enable data exfiltration and intelligence gathering.
  • Order also claims models themselves “may contain malware or vulnerabilities,” citing prior research on DeepSeek’s mobile app.
  • Thread references:
    • Work showing models can be backdoored or embed hidden capabilities in weights.
    • Demonstrations of trigger phrases causing malicious tool calls or insecure code.
    • Reports that DeepSeek produced more insecure code under certain political contexts (disputed as either sabotage or an artifact of heavy-handed “alignment”/censorship).
  • Many argue open weights are like opaque binaries: downloadable but not auditable in practice; you can’t reliably detect backdoors from raw numbers.

Comparison with U.S. Models and Geopolitics

  • Some say the same logic could justify other countries banning U.S. models (CLOUD Act analogies, surveillance capabilities).
  • One camp: this is prudent statecraft—don’t let an adversary’s opaque AI into government systems.
  • Another camp: this is largely protectionism/regulatory capture to shield U.S. frontier labs whose closed APIs are at risk from cheaper, strong Chinese open‑weight models.

Open Weights, Trust, and Ecosystem Impact

  • Strong disagreement over whether open weights are safer:
    • Pro: self‑hosting avoids foreign data collection; you control environment and network.
    • Con: training pipeline is opaque, hidden behaviors can persist, and open vs closed makes little difference to deep backdoors.
  • Concern that banning open‑weight Chinese models from federal use will chill open‑weight releases generally and push agencies toward domestic proprietary APIs.

Practical and Downstream Effects

  • Questions about:
    • Whether services like OpenRouter that proxy DeepSeek are now unusable for federal work.
    • Whether code or libraries written with DeepSeek count as “developed using” it (widely seen as unenforceable but chilling).
  • Some note other governments (e.g., Australia) have already banned DeepSeek, and expect more bans on other Chinese models (Qwen, GLM, etc.) as they grow.