OpenAI Bans Russia and Iran-Linked Accounts Over Secret Influence Operations

OpenAI has suspended Russian- and Iran-originated ChatGPT accounts that are associated with covert influence operations among audiences in multiple countries. Both groups used AI along with traditional methods, such as fake accounts, media pitches, and invented content.

Russian Network Targeted Latin America Through a Fake Research Group

The ChatGPT-based Russian operation was directed at audiences in Latin America, with a significant portion targeting Ukraine. Some material also covered domestic politics in Argentina, Bolivia and Ecuador, according to OpenAI.

The network allegedly controlled the Social Research Center through a fake persona called “Mia Clark.” Local employees were apparently not aware they were employed by Russian operators and did research in good faith.

The use of ChatGPT was primarily for internal reports, with some requests to create fake letters and audio scripts, and a few for Spanish content localized in a specific region. The campaign was rated as a Category 5 disrupted influence operation by OpenAI, the highest rating it’s ever assigned.

The case comes at a time when AI security is being increasingly looked at after an AI agent from OpenAI gained access to Australian government records in internal testing.

Iran-Linked Accounts Adopted Journalist Personas

OpenAI also blocked Iranian-origin accounts that used seven fake journalist identities to pitch articles to online publications. The operation, dubbed “Bogus Bylines,” primarily targeted the U.S.-Iran conflict and produced content in English and Persian.

The company found nearly 100 articles, both published and syndicated, in about a dozen Internet sites using those bogus names. A few personas also had social media accounts which helped to spread published articles and related messaging.

The posts about Iran, the United States and Israel elicited comments, but not much engagement. OpenAI gave its assessment of the article-publishing effort a Category 4 rating and a Category 2 rating for the social media operation.

The results further come on the heels of broader calls from executives in the AI space for slowing down frontier model development and independent safety testing.

Published Articles Gave Campaigns a Wider Reach

The two operations distinguished themselves by getting content to appear beyond their own social media platforms. OpenAI observed that campaigns that managed to gain coverage or publication via traditional channels tended to reach larger audiences than campaigns that were limited to social channels.

This is the trend that has cropped up in the 30 influence operations that OpenAI has studied since early 2024. With the help of AI, tasks like drafting, editing, translation, and internal reporting were faster, but these campaigns still required the use of outdated influence tactics for genuine people.

The results are also relevant to the growing concerns about the use of sophisticated AI tools, which are not just aimed at influencing elections. As we reported, security researchers used Anthropic’s Claude while testing vulnerabilities affecting ChatGPT and Codex accounts. The final research process revealed vulnerabilities that could allow access to connected services and internal resources.