China AI firms accused of mass model distillation
CISA, NSA, and FBI say DeepSeek, Alibaba, Moonshot AI and others systematically extracted proprietary data from U.S. frontier AI models via industrial-scale distillation.
Sourcecisa.gov — open the original
What to check
- Applies to
- Whether this concerns you is not assessed; the names the report mentions are listed below.
NSA, CISA, and FBI issued a joint advisory alleging that multiple China-based AI companies—DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI—have conducted industrial-scale knowledge distillation campaigns against U.S. frontier AI models (Claude, GPT, Gemini, Grok variants) since at least late 2024. The advisory characterizes this activity as extending beyond legitimate distillation research into systematic extraction of proprietary functionality, reasoning capabilities, and domain-specific optimizations, violating U.S. AI companies' terms of use. The agencies assess this is likely conducted with Chinese government awareness and represents a core strategy—not a supplement—for closing the capability gap with U.S. models while reducing training costs and timelines.
The advisory details TTPs mapped to the MITRE ATLAS framework plus four novel techniques not yet catalogued: regional restriction/subscription evasion via bulk premium account procurement, centralized request-routing infrastructure spanning native APIs/cloud providers/third-party aggregators/"transfer station" proxies, automated metadata sanitization to strip organizational identifiers, and systematic quota/cost optimization. Detection indicators include shared accounts across multiple IPs/user agents, 24/7 sustained usage without human variation, anomalous subscription-to-API usage ratios, and new accounts hitting maximum usage immediately rather than ramping gradually.
Recommended mitigations for U.S. AI companies include behavioral detection and monitoring of subscription/usage patterns, targeted response degradation (differential privacy, downgraded model responses) for suspected malicious distillation traffic without alerting the requester, and cross-organization intelligence sharing of infrastructure and behavioral indicators to reveal distributed campaigns that appear isolated to any single provider. The advisory also references NIST AI 100-2e2025 mitigations such as differential privacy, pre/post-training interventions, and prompt formatting controls.
Mentioned in this report
Source reporting: https://www.cisa.gov/news-events/cybersecurity-advisories/aa26-251a
Free public brief. The indicators are in the app with a free account, no card. Slack alerts on the vendors you run come with the paid plans, and a new account gets three days of them.
What this brief leaves out
This page is the free tier: the write-up, the severity, the names. In the app the same report carries its extracted indicators and live profiles of the campaigns it names. The paid plans add the security research and news reporting the free tier leaves out.
It sits in a corpus of 16,445 reports from 149 sources, 666 of them written in the last seven days, and it grows through the day.
A new account starts with three days of all of it, no card, then keeps the government and CERT reporting free. Slack alerts are a paid-plan feature: name the vendors you run and Slack hears about the reports that touch them, not the rest of the week's. Nothing is charged unless you subscribe.
Create a free account What it costs