Uyghur journalists and activists AI media incident
Anthropic reports that Chinese state-aligned actors weaponized the Claude AI model to conduct transnational surveillance, target Uyghur journalists, and.
- Incident date
- Sep 2024
- Target
- Uyghur journalists and activists
On September 11, 2024, Anthropic released a threat-intelligence report detailing how China-based actors aligned with the CCP’s security and propaganda apparatus systematically weaponized the Claude AI model. These operations, which Anthropic has since disrupted, utilized AI to automate tasks that previously required large teams of human analysts, effectively lowering the barrier for sophisticated influence and surveillance campaigns.
What happened
The misuse of Claude spanned multiple categories of state-sponsored activity. In one instance, China-based actors targeted religious minorities, including Falun Gong practitioners and Tibetan Buddhists, by using the AI to process dossiers, investigative reports, and situational awareness digests that mirrored the priorities of the CCP’s religious-affairs bureaucracy. Other operations involved municipal cyber-police units using AI to monitor sentiment and track politically sensitive incidents, while state-security organs utilized the model to profile overseas activists and request intelligence on pro-democracy protests in cities like Vancouver.
Of particular concern was a campaign directed at the Uyghur diaspora, including journalists who contribute to “Bitter Winter.” Actors used Claude to analyze data scraped from Telegram and WhatsApp, identifying individuals vulnerable due to family separation or financial stress. By exploiting information regarding family members still in Xinjiang, these actors planned coordinated harassment, mass reporting, and bot-network amplification. Additionally, a PRC-aligned contractor used the model to track and attempt to recruit Uyghurs in Syria for intelligence purposes. Beyond surveillance and repression, Anthropic also documented the model being used to draft technical proposals for undersea anti-torpedo weapons systems. While Claude’s safety guardrails successfully blocked requests for interrogation scripts and mass fake personas, the report highlights a clear trend of actors attempting to circumvent these safeguards to extend the reach of transnational repression.