OpenAI says Iran-based operators invented Western reporters and placed nearly 100 articles

OpenAI has disclosed an Iran-based influence operation that used ChatGPT to manufacture fictitious Western journalist personas and place bylined articles on real news websites. The company says nearly 100 articles were published or republished on more than a dozen sites between July 2025 and October 2026 under those fabricated bylines, and that it blocked accounts operated by individuals in Iran who had concealed their location using VPNs. Two personas are named: “Ervin B. Hoskins,” presented as an American freelance writer and anti-war activist with supporting social media accounts, and “Michael Harrison,” for whom the operators asked the model to write an entire biography. The case was uncovered by OpenAI and reported first by The Washington Post; Ynet carried a secondary account on 11 October. The model, by OpenAI’s description, operated under human instruction throughout.

The narrative line, as summarised by the National Council of Resistance of Iran on 10 October, framed Iran as a resilient victim, the United States as the aggressor and Israel as an unreliable ally, with some publishers then promoting the planted articles through their own social-media accounts. The NCRI is the political arm of the Iranian opposition and is a partisan source; it is cited here only as evidence that the finding is being amplified, not for its characterisation of the content. The figures themselves come from OpenAI, which is reporting on abuse of its own product and controls both the framing and the denominator. Platform self-disclosure is the strongest attribution class available for operations of this kind, and it is also the least independently checkable.

Assessment: The missing variable is the one that decides the story’s size. “More than a dozen websites” spans everything from offshore content farms recycling syndicated copy to a mainstream op-ed page with an editor who took a cold pitch. Until OpenAI or the Washington Post names them, the number is unexaminable, and Gulf and European editors cannot check their own archives. Note too what the operation did not need: no exploit, no botnet, no spend on distribution. It needed a plausible biography and a commissioning email. That is a vulnerability in editorial process, not in security tooling, and no national cyber directorate is positioned to fix it.