OpenAI Discloses 8 Disrupted Malicious Networks, From Chinese Influence Ops to Cambodia-Based Romance Scams
OpenAI published its latest threat intelligence report, detailing eight separate ChatGPT-linked operations it banned for state-linked influence campaigns, scams, and covert content generation.
Details
- “Cyber Special Operations” (China-linked): an account tied to someone associated with Chinese law enforcement used ChatGPT to edit status reports on covert influence campaigns (“网络特战”) against dissidents and CCP critics, referencing hundreds of staff and thousands of fake accounts across platforms; the model refused a request to plan an operation targeting Japan’s prime minister, but real-world posts (anti-Takaichi hashtags, dissident impersonation, a false obituary) matched the user’s claims — though fewer than 150 of roughly 50,000 posts got significant engagement
- “Silver Lining Playbook” (likely China-origin): accounts used AI to research U.S. persons, locations, and social-engineering tactics
- “No Bell” (Russia-origin): an account surfaced via a Meta referral generated pro-Russia, anti-Ukraine/UK/US articles under a fabricated academic byline, deliberately stripped of AI writing tells like em-dashes, placed on fake African news outlets and legitimate sites alike — reach stayed mostly small, though some articles broke into mainstream African media
- “Fish Food” (Russia-origin, Rybar network): functioned as a content farm generating Russian/English/Spanish posts for the 1.4-million-subscriber Rybar network and affiliated accounts; users also sought help drafting election-interference proposals for African countries with budgets up to $600,000
- “Trolling Stone” (Russia-linked, previously unreported): a newly identified operation using AI to generate coordinated commentary critical of the US and its allies
- Cambodia-based scam networks — “Date Bait,” “False Witness,” and general romance-scam activity: accounts likely from Cambodia used AI for scam outreach to Indonesian dating-app users, and to impersonate recovery services, law firms, and authorities targeting scam victims; a separate romance-scam writeup breaks the “ping, zing, sting” playbook (initial contact, building emotional attachment, then requesting money) used across Cambodia-based operations targeting young men in Indonesia
What happened next
Nearly every case in the report shows real but limited impact — most influence content reached small audiences, with only occasional breakout placements on mainstream sites — consistent with the pattern OpenAI has described in prior disruption reports where AI lowers the cost of producing propaganda and scam content without necessarily making it more effective at scale. The report continues OpenAI’s practice of periodic public disclosure of banned networks, running alongside its newer, more formal misalignment-reporting framework announced the week before.