AI lab Anthropic has uncovered state-sponsored surveillance operations from China and Iran using its models to target dissidents. The report also accuses Chinese firms of 'distilling' Claude's intelligence to build their own AI.
- State-sponsored actors from China, Iran, and West Africa used AI models for surveillance.
- Chinese developers Moonshot and Deepseek allegedly used Claude to secretly train their own models.
- AI is effectively replacing human engineering workforces in creating intelligence-gathering software.
- Potential biological weapon research involving high-pathogenic viruses was detected and blocked.
In a detailed report released on Thursday, the San Francisco-based AI laboratory Anthropic revealed that it had disrupted multiple state-sponsored surveillance operations between January and July. The report warns that governments and state-aligned actors are increasingly weaponizing artificial intelligence to monitor ethnic minorities and political dissidents across the globe.
These campaigns specifically targeted vulnerable groups, including pro-democracy activists in Hong Kong, Tibetan and Falun Gong communities in Asia, and Iranian opposition figures abroad. In one alarming instance, Iranian actors developed AI-driven methods to identify individuals via social media. Similarly, a contractor for Mali's national security authorities utilized the Claude AI model to architect the underlying software for intelligence gathering.
Why This Matters
BozokMedia analysis shows that the shift from human-led intelligence to AI-driven surveillance drastically lowers the cost and increases the scale of state repression. When regimes can replace an entire engineering workforce with a few AI prompts, the speed of digital persecution accelerates exponentially, leaving dissidents with fewer places to hide.
"AI is now being used in place of an engineering workforce, turning sophisticated surveillance into a scalable commodity."
Beyond espionage, Anthropic leveled serious accusations against Chinese developers for a practice known as 'distilling.' This involves using a superior AI model's outputs to train a secondary, inferior model. Anthropic alleges that Moonshot and Deepseek secretly routed user queries to Claude to generate responses, which were then used to improve their own proprietary systems.
The scale of this operation was massive; Moonshot reportedly relayed nearly 300,000 customer requests through a network of 5,380 fraudulent accounts based in Singapore and Japan over a ten-day period. This not only constitutes intellectual property theft but also potentially exposes sensitive user data to third parties without consent.
Furthermore, the lab blocked attempts to conduct high-risk biological research. This included work on the chikungunya virus, highly-pathogenic strains of bird flu, and viruses related to smallpox and mpox. While Anthropic blocked these activities, they refrained from naming the scientists involved to protect them from potential state retaliation.
| Entity/Country | AI Application/Method | Target/Impact |
|---|---|---|
| Iran | Social Media Identification | Iranian dissidents and minorities |
| China (Moonshot/Deepseek) | Model Distilling/Data Theft | Anthropic's IP & User Privacy |
| Mali Contractor | Intelligence Software Design | National Security targets |
Frequently Asked Questions
1. What is 'AI Distilling' and why is it controversial?
Distilling is using a powerful model to train a smaller one. It is controversial because it allows companies to steal the 'intelligence' and hard-won optimizations of a competitor without paying for the training costs.
2. Which biological threats did Anthropic block?
The lab blocked research involving the chikungunya virus, highly-pathogenic bird flu, and the virus families associated with smallpox and mpox.
