Two operations, one older playbook
OpenAI's October 8 report describes two covert influence operations that looked modern at the surface and familiar underneath. Both used ChatGPT. Neither depended on a novel model capability to work.
The Russia-origin cluster, which OpenAI calls Dark Clark, targeted Latin America. It appears to have controlled a self-described research platform called the Social Research Center through a fabricated persona, Mia Clark. Internal reports discussed hiring, firing, pay and research projects. The people doing work on the ground appear not to have known that a Russian operation sat behind the organization.
The Iran-origin cluster, named Bogus Bylines, took a newsroom route. Seven invented Western journalist identities pitched English-language articles, mostly about the US-Iran conflict, to small and medium international-affairs outlets. OpenAI identified almost 100 articles published or syndicated across roughly a dozen publications.
Those are not chatbot tricks. They are institutional tricks. The model made parts of the workflow cheaper and more fluent, but the operation still needed cover stories, plausible identities, managers, distribution and people willing to treat the front as real.
Dark Clark used ChatGPT to report upward
The Russian case is the cleaner rebuttal to the idea that generative AI was the engine. OpenAI says the operators' main use of ChatGPT was writing and updating internal reports for an unknown superior. In most cases, the company did not observe its models creating the campaign material itself.
Some of those reports described genuine activity. OpenAI found public traces matching several claimed operations, including fabricated material that triggered fact checks or official denials in Ecuador and Bolivia. A scheme involving fake instructions to schools in Peru also appears to have generated press coverage and political comment in Europe.
Other claims were inflated. The operators sometimes tried to take credit for developments they did not cause. That matters because threat actors' own performance reports are not neutral evidence; they can be propaganda aimed inward at whoever pays the bills.
OpenAI nevertheless placed Dark Clark in Category 5 on the Breakout Scale, the first operation it has disrupted at that level. The score reflects visible reactions from politicians in multiple countries, not proof that every claim in the operators' reports was true.
Bogus Bylines found the real bottleneck
The Iranian operation used ChatGPT more directly in publication. Operators supplied English drafts, asked the model to compare them with a specific outlet's submission rules, refined the copy, then generated pitch emails. The seven bylines came with biographies and, in some cases, social profiles that made them look like ordinary freelance journalists.
The striking number is not how many comments the model could produce. It is how many editors accepted the articles.
OpenAI found almost 100 published or syndicated pieces from July 2025 through October 2026. Some were promoted by the outlets' own social accounts, giving the campaign access to audiences it could not easily build itself. That work reached Category 4.
The batch-generated comments went nowhere. OpenAI says none drew substantial likes, shares or reply chains, and rates that channel Category 2. The operators even counted views on the posts they replied to rather than views of their own replies, a method that made weak performance look stronger.
The contrast is the report's most useful result. Cheap synthetic participation did little. Borrowed editorial legitimacy did much more.
Detection focused on prose will miss the operation
A detector can flag awkward phrasing, repeated syntax or likely machine-generated text. None of that answers the important questions here: Is this journalist real? Who controls the research center? Why do its employees not know the ultimate sponsor? How did a forged document move from a private workflow into a credible publication?
The defence has to sit closer to the point where trust is granted. Editors need identity checks that go beyond a polished biography and active social profile. Research organizations need transparent ownership and funding. Newsrooms accepting contributed articles need a record of who introduced the writer, how credentials were verified and whether several supposed freelancers share infrastructure or story themes.
That work is slower than running copy through an AI detector. It is also aimed at the mechanism that produced reach.
The AI label can obscure the lesson
OpenAI has now exposed 30 covert influence operations over two and a half years. Its own comparison shows the highest-reach cases clustering around external publication rather than fake social distribution. That is an uncomfortable finding for platforms because it shifts attention away from model output and toward the ordinary institutions that confer credibility.
AI increased efficiency, linguistic fluency and editorial polish. It also left evidence inside the service that helped OpenAI connect accounts and reconstruct workflows. But the report does not describe autonomous persuasion at machine speed. It describes operators using a general-purpose tool inside organizations built to deceive other organizations.
The response should follow the evidence. Keep improving model-abuse detection and account disruption, but do not mistake the tool for the campaign. The breakthrough came from false fronts passing as real ones, and from real publications opening the door.
Sources
- [1] OpenAI, “Disrupting AI-enabled ‘false front’ operations” (October 8, 2026)Read source
- [2] The Decoder, “OpenAI uncovers Russian and Iranian influence ops that planted fake stories in real news outlets” (October 9, 2026)Read source
- [3] Qazinform, “OpenAI disrupts Russian, Iranian AI-enabled influence operations”Read source