OpenAI has banned a ChatGPT account tied to Chinese law enforcement after the user attempted to leverage the model for covert influence operations. The most striking attempt involved planning a disinformation campaign targeting Sanae Takaichi, Japan’s first female prime minister. The user sought assistance in crafting a multi-pronged strategy that included amplifying negative comments, fabricating complaints from fake immigrant accounts, and accusing Takaichi of far-right leanings. While OpenAI’s safety systems refused to generate the initial attack plan, the user later returned to polish status reports on the same operation, suggesting the campaign proceeded using other tools.
This incident exposes a broader, resource-intensive strategy termed “cyber special operations” by the actors themselves. Evidence indicates this effort involves hundreds of staff and thousands of fake accounts across dozens of platforms. The scope is global, targeting not only domestic dissenters but also critics of the Chinese Communist Party abroad and foreign officials. Tactics range from abusive reporting of social media accounts to forging documents and impersonating US officials. The sheer scale suggests a well-funded, sustained apparatus rather than isolated rogue actors.
Crucially, the investigation reveals a shift in tooling. While the actors tried to use ChatGPT for editing and polishing reports, their operational heavy lifting appears to rely on locally deployed Chinese open-weight models like DeepSeek and Qwen. OpenAI’s open-source investigative techniques linked the user’s activity to known influence networks, including the “Spamouflage” campaign previously attributed to Chinese law enforcement by Meta. A specific website, revealscum.com, served as a digital fingerprint connecting these recent attempts to earlier disinformation infrastructure.
The refusal of Western AI models to assist in direct planning may be pushing these operations toward fully domestic AI stacks. This creates a fragmented technological landscape where state-sponsored influence campaigns are increasingly insulated from Western safety guardrails. For policymakers, the challenge is no longer just about detecting fake content, but understanding how sovereign AI ecosystems are being weaponized for diplomatic sabotage. The target list now includes heads of state, raising the stakes for international cyber norms.
