US Accuses Six Chinese AI Firms of Stealing Frontier Model Capabilities
Chinese firms allegedly copied US AI models to cut development costs and time.
1Accusations Against Chinese AI Firms
The US government has accused six Chinese AI firms—DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI—of conducting industrial-scale attacks to extract capabilities from US frontier AI models like Claude, GPT, Gemini, and Grok. These actions, allegedly carried out with Chinese government awareness, have allowed these firms to significantly shorten development timelines and reduce financial expenditures. The attacks involve exploiting model inference APIs and using bulk-bought fake accounts to execute coordinated queries.
2Methods of Model Extraction
Chinese firms allegedly use prompt injection techniques to jailbreak models, forcing them to reveal hidden reasoning steps. For example, DeepSeek crafted prompts instructing models to articulate internal reasoning step-by-step. Another method involves bulk-buying fraudulent accounts to send thousands to millions of similar queries, far exceeding legitimate usage. These accounts often rely on proxy gray markets to evade geographical restrictions and route requests through multiple pathways.
3US Recommendations for Mitigation
US agencies recommend AI firms improve detection of sophisticated campaigns by monitoring anomalous prompts, accounts, and behaviors. Suspicious indicators include bulk premium subscriptions shared across developer teams and new accounts hitting maximum usage immediately. Firms should also strengthen identity verification and track enterprise subscription usage. These steps aim to flag bulk deployments with pre-engineered templates, though they may raise privacy concerns for legitimate users.
4Dumbing Down Model Responses
To counter distillation attacks, US agencies suggest subtly altering responses—such as presenting correct information with different reasoning or reducing depth—to decrease payoff for attackers. Firms could also secretly switch malicious accounts to inferior models without notice. However, this poses technical challenges, as Chinese firms employ adaptive discovery to detect and switch to smarter models within 24 hours, using automated QA systems to identify degraded outputs.
5Risks to Legitimate Users
Mitigation efforts risk collateral damage: legitimate users might be switched to dumber models without warning or experience withheld capabilities. Adding noise to outputs could restrict further queries, frustrating users. OpenAI faced backlash last year when its routing system defaulted to less capable variants unless prompted explicitly. Agencies advise balancing security with user experience, accepting trade-offs like lower precision to prevent capability theft.
6Long-Term Defense Strategy
Agencies emphasize the need for AI firms and allied governments to coordinate long-term defenses against industrial-scale theft. This includes sharing threat intelligence and developing resilient model architectures. The goal is to maintain US leadership in AI while mitigating risks from adversarial exploitation. However, overly aggressive measures could harm innovation and user trust, requiring careful calibration of security and usability.
7Impact on AI Development Race
The allegations highlight the intensifying AI race between the US and China, where intellectual property theft could erode US competitive advantages. By copying frontier models, Chinese firms bypass years of R&D, accelerating their progress. This threatens to narrow the gap in AI capabilities, prompting calls for stricter safeguards and international cooperation to protect proprietary technologies.
8Broader Implications for AI Governance
The incident underscores the need for robust AI governance frameworks to address industrial espionage while fostering innovation. Policymakers must balance openness with security, ensuring that defensive measures do not stifle collaboration or transparency. As AI becomes a strategic asset, international norms and agreements may be necessary to prevent destabilizing theft and maintain fair competition in the global AI ecosystem.