NSA, CISA and FBI accuse six Chinese AI firms of industrial-scale model theft

[FrostDrifter]

The NSA, CISA and FBI issued a joint cybersecurity advisory naming DeepSeek, Alibaba, Moonshot AI, MiniMax, StepFun and Z.AI for conducting what the agencies call industrial-scale knowledge distillation campaigns against US AI companies, extracting billions of tokens across millions of requests from models including Claude, GPT, Gemini and Grok since at least late 2024. The advisory states plainly that while distillation is a legitimate research technique, these companies are engaging in aggressive, malicious and targeted extraction of restricted proprietary functionality at a scale the agencies say forms the core of their AI development strategy rather than merely supplementing it

The agencies describe the activity as likely occurring with Chinese government awareness, though the advisory doesn't present direct evidence for that specific claim and, notably, never uses words like theft, illegal or crime anywhere in its roughly 3,600 words. Instead, it frames the issue as a violation of the targeted companies' terms of service, and provides API providers with behavioral indicators to watch for, including accounts running 24/7 without normal human variation, new subscriptions immediately hitting maximum throughput, and shared accounts accessed from multiple IPs

Unusually, the advisory recommends that providers avoid informing suspected distillers when their access gets downgraded, reasoning that alerting them would let them adapt their evasion techniques and know when to roll back training runs. The advisory doesn't announce any specific enforcement action, sanction or legal proceeding, leaving open whether it represents an opening move toward something like export controls or civil litigation, or stands alone as a warning. Curious what people think about behavioral usage patterns being the primary evidence here rather than any technical or forensic proof, is that a reasonable basis for a formal government advisory naming specific companies, or does it set an uncomfortably low bar

Aa26 251Awww.cisa.gov

[Ellie85]

Claude, GPT, Gemini and Grok all being named as targeted models shows this is being treated as a shared industry problem across competing US labs, not just one company's individual grievance with a specific rival

[Harbour17]

Behavioral indicators like 24/7 usage without human variation describe basically any legitimate enterprise AI agent deployment too, that detection profile risks flagging plenty of genuine automated customers alongside actual bad actors

[AlexaBliss_Fan]

No accompanying enforcement action or sanction makes this feel like an opening move rather than a conclusion, worth watching whether an export control action or civil suit actually follows and cites this advisory as its foundation

[CMPunk_Mike]

Likely with Chinese government awareness doing a lot of work without any actual evidence cited behind it shows how much of this advisory rests on inference and pattern rather than confirmed intelligence, worth reading that specific claim cautiously

[CosmicRay91]

Recommending providers stay silent about downgrading suspected accounts is a reasonable operational security tactic on its own, though it also means affected customers never get a clear explanation for degraded service if they're wrongly flagged

[Louise84]

Never using the words theft, illegal or crime anywhere in the entire advisory is a telling detail, that's a carefully lawyered document making a serious sounding accusation while stopping well short of any formal legal claim