AI agents are turning influence campaigns into self-running software

The next phase of online propaganda is not better-written posts. It is software that creates the accounts, invents the identities and manages the campaign.

Iranian and Chinese-linked operators have used hundreds of AI agents to build fake social-media networks, publish political material and coordinate activity across Instagram, Facebook, X and TikTok, according to US officials and security researchers cited by The New York Times.¹

The agents were built on Chinese open-source models and required limited human supervision. Operators supplied the political objectives; the software handled much of the repetitive work involved in creating and running the campaign.

• AI agents reportedly managed several stages of the operation, from creating fake accounts to coordinating posts and interactions across platforms.¹
• The campaigns were linked by US officials and researchers to actors in Iran and China, with Israeli firms also reported to be developing similar systems.¹
• Anthropic found operators using Claude to produce fake reporters, political spokespeople, profile photographs, forged documents and persistent persona systems.²
• The technology is not yet a guaranteed route to mass influence: Anthropic said most operations it detected attracted little genuine engagement or were stopped before building an audience.²
• There is no public evidence that the campaigns described changed public opinion or produced a measurable political result.¹²

That limitation matters. The immediate danger is not that an AI swarm can magically persuade millions of people. It is that the cost of trying has fallen sharply. A campaign that once required writers, account managers, translators and supervisors can increasingly be expressed as a set of instructions and left running.

Anthropic’s September report shows how far this is moving beyond text generation. Operators used Claude to prepare doctrine manuals, target databases, opposition dossiers and systems for fake personas. They stored rules, approved sources and evasion tactics in persistent files, allowing separate agents to work from a common playbook without direct human coordination.²

The platforms still detected and removed some of the activity. Anthropic said it banned the accounts it identified, while Meta separately removed an Iran-linked network targeting American politicians and journalists.²³ But the public record does not yet show how widely the newer campaigns travelled, whether any escaped platform enforcement or whether they shifted real-world opinion.

That is the uncomfortable distinction now being blurred by the word “autonomous”. These systems are not independent political actors. Humans still choose the cause, targets and desired outcome. But the human role is moving upwards — from writing each message to commissioning an operation that can keep producing, testing and adapting them.

The result is a more dangerous information environment even when individual campaigns fail. More actors can cheaply manufacture the appearance of a crowd, and more citizens will have to treat apparently ordinary political activity as potentially synthetic. The burden of proving that a voice is real is shifting from the operator to the public.

Footnotes

1. Sheera Frenkel, Eli Tan and Dustin Volz, “Iran and China Create First-of-Their-Kind Autonomous A.I. Influence Campaigns”, The New York Times, 18 September 2026: https://www.nytimes.com/2026/09/18/technology/iran-china-autonomous-ai-influence-campaigns.html ↩ ↩² ↩³ ↩⁴

2. Anthropic, “Detecting and countering misuse of AI: September 2026”, September 2026: https://www.anthropic.com/threat-intelligence-report-september-2026 ↩ ↩² ↩³ ↩⁴ ↩⁵

3. Sara Fischer, “Exclusive: Meta disrupts Iran-linked AI operation targeting politicians, journalists”, Axios, 27 August 2026: https://www.axios.com/2026/08/27/facebook-instagram-iran-ai-disinformation ↩