From Ars Technica · · 2 min
US Accuses Six Chinese AI Firms of Mass Distillation Attacks
Federal agencies claim Chinese tech companies are using fake accounts and prompt injections to steal capabilities from leading US AI models.
In brief
Federal agencies claim Chinese tech companies are using fake accounts and prompt injections to steal capabilities from leading US AI models. US intelligence agencies want tech companies to join forces and aggressively monitor accounts to stop Chinese firms from distilling US frontier AI models. Originally reported by…
US accuses Chinese AI firms
The NSA, CISA, and FBI released a joint statement accusing six Chinese AI companies of conducting industrial-scale attacks to copy US frontier AI models.
“China-based AI companies that conduct industrial-scale distillation against US AI models see significantly shorter AI development timelines”
Swarms of fake accounts execute queries
Attackers bulk-buy fraudulent accounts and route requests through proxy networks. They run thousands to millions of coordinated queries featuring identical prompts to extract model capabilities.
Prompt injection forces step-by-step reasoning
Chinese developers use prompt injection techniques to jailbreak US models. DeepSeek instructed models to articulate their internal reasoning step by step.
“DeepSeek employed prompts instructing models to imagine and articulate the internal reasoning behind completed responses”
Government urges aggressive account monitoring
Federal agencies recommend that US AI companies flag suspicious accounts hitting maximum usage immediately. They also suggest tracking subscription usage ratios and strengthening identity checks.
Dumbing down model responses defensively
Agencies suggest subtly altering outputs for flagged accounts, such as reducing reasoning depth or presenting different logic. US firms could secretly switch malicious users to inferior models without warning.
Legitimate users risk collateral damage
Secretly downgrading accounts could inadvertently hurt real customers. Legitimate users might experience degraded output quality, shorter answers, or restricted follow-up queries.
Attackers quickly adapt to degraded outputs
Chinese companies run automated quality assurance to distinguish service glitches from deliberate data degradation. Some automated systems detect smarter models and switch targets within 24 hours.
Targeted models and stolen capabilities
The report highlights DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI. Targeted capabilities include software engineering, math, writing optimization, and agentic functions.
Beijing calls US claims groundless
A Chinese Ministry of Foreign Affairs spokesperson dismissed the accusations, attributing China's AI growth to technological self-reliance. Officials accused the US of running a smear campaign.
Cross-border AI ties and political pressure
Chinese officials noted that US startups also rely on affordable Chinese models. The dispute comes as China expands its computing power and leaders prepare for bilateral talks.
Here's the gist
US intelligence agencies want tech companies to join forces and aggressively monitor accounts to stop Chinese firms from distilling US frontier AI models.
