OpenAI Banned a Russian Operation and an Iranian One That Invented Seven Journalists -- Its First 'Category 5' Takedown

OpenAI said Thursday it has banned two unrelated covert influence operations that used ChatGPT -- one from Russia, one from Iran -- in what the company calls its first-ever Category 5 case on the six-point (1-to-6) scale it uses to rate how far a propaganda operation actually broke out into real audiences. The Russia-origin operation, which OpenAI calls Dark Clark, built a fake Latin American think tank around an invented persona; the Iran-origin operation, Bogus Bylines, invented seven Western journalists and used them to place almost 100 articles at real outlets. OpenAI says it has now disrupted 30 such operations since it began publishing these reports in May 2024. (Thirty operations across two and a half years works out to roughly one disclosed every month -- a reporting rhythm, not a rate of occurrence, since the count only includes operations OpenAI itself detected and chose to disclose.)

[The operations used their entities] to launder geopolitical, conflict-related messaging into their target audiences.

Dark Clark's center of gravity was something called the Social Research Center, a fabricated think tank fronted by a persona OpenAI calls "Mia Clark." The operators, mostly writing in Russian from behind VPNs (OpenAI blocks direct access from Russia), used ChatGPT to draft internal reports on three workstreams: denigrating Ukraine and discouraging enlistment in its armed forces, interfering in Bolivian and Argentine domestic politics, and running the Center's own front, which posted more than 60 articles on its own site. Per OpenAI's account, the operation's claimed handiwork includes a fabricated 2024 rumor that Argentine President Javier Milei bought jeweled dog collars, a May 2026 fake email impersonating a Lima education office, fabricated audio in March 2026 purporting to be Ukraine's honorary consul in Ecuador, and fabricated audio in May 2026 of a Bolivian utility worker claiming a water-service emergency. OpenAI is explicit that these are the operation's own self-reported claims of impact, not independently confirmed outcomes.

The one that borrowed a newsroom's byline list

The Iran-origin operation worked differently: rather than impersonating officials, it invented seven complete people -- among them Ervin B. Hoskins, Noah Lamington and Sophia Gonzalez -- and used ChatGPT to draft, polish and pitch opinion and analysis pieces under their names to roughly a dozen small and mid-sized outlets covering international affairs. OpenAI says close to 100 articles ran this way between July 2025 and this month, with the pace picking up after the U.S.-Iran conflict began. The operation also ran a smaller, separately rated commenting campaign -- more than two dozen batches of hostile replies to the UK-based outlet Iran International -- that OpenAI scored far lower (Category 2) than the article-planting workstream (Category 4), because engagement on the comments was thin and much of what little there was traced back to the operation's own accounts. OpenAI says it couldn't identify who was actually behind the operation, beyond noting it looked more like a commercial, for-hire campaign than a state messaging shop.

The closest thing either operation has to outside confirmation isn't a shared data set -- it's a coincidence of enforcement. OpenAI's report notes that Meta, independent of anything OpenAI found, disrupted a persona named "Michael Harrison" in March 2026 and banned a related Instagram account in August, months before OpenAI's own ban. Two different companies' moderation teams independently running into the same fake person is a stronger signal than either company's account on its own.

The ratings in that table are OpenAI's own, applied by OpenAI to findings OpenAI itself collected -- which raises the obvious next question: checked against what, and by whom?

These disclosures are voluntary. Nothing in U.S. law requires OpenAI, Anthropic or any other model maker to publish an account of influence operations it finds running on its own platform -- the practice exists because OpenAI chose to start it, in May 2024, well ahead of any regulatory requirement to do so. Social platforms like Meta's face disclosure expectations for coordinated inauthentic behavior under the EU's Digital Services Act regime; a chatbot whose output gets published somewhere else entirely does not, yet, face an equivalent. That gap is why Dark Clark and Bogus Bylines surface in a company blog post rather than a regulatory filing: a rule that extended platform-style disclosure duties to the companies whose models generate this content, not just the platforms that host it, would turn today's voluntary report into a compliance requirement.

What's established, and what's just OpenAI's own rating

The strongest fact either report actually establishes is the ban itself -- the accounts are gone, the output is attributed, and in Bogus Bylines' case, a specific, checkable list of fake names now exists for the affected outlets to check against their own archives. Everything about how far either operation's propaganda actually reached a real audience, rather than just a handful of pitched editors and an operation's own comment threads, still runs on OpenAI's account of its own findings.