Iran and China-Linked Campaigns Use AI Agents to Run Fake Accounts

The new capability lowers the labor required to create and coordinate online personas. Reach is becoming cheaper to manufacture; persuasion still has to be earned.

Iranian and Chinese influence operations have used AI agents to create and operate networks of fake social-media accounts with limited human intervention, according to reporting based on U.S. officials and security researchers. The systems combined downloadable Chinese AI models with agents able to open accounts, generate posts and coordinate activity across Instagram, Facebook, X and TikTok.

One Iranian operation targeting U.S. audiences accumulated nearly 80,000 followers in the first half of 2026. Public detail on the Chinese campaign is thinner, so the two operations should not be treated as equal in demonstrated scale.

Autonomy compresses the labor budget

Traditional influence operations require operators to maintain identities, write material, monitor events and coordinate posting. Agents can automate several of those tasks and keep more personas active at once.

Lower labor intensity changes the cost curve even when the content is mediocre. A campaign can test more identities, react faster to news and sustain activity across more platforms without adding a human operator for every account.

Anthropic's separate September threat-intelligence work shows the same direction of travel in other influence activity, including AI-assisted impersonation and state-linked uses of models. It also warns that high output does not guarantee authentic engagement.

The distinction is central. Nearly 80,000 followers demonstrate distribution, not persuasion. Platforms and advertisers still need measures of organic engagement, downstream sharing and real-user conversion before equating automated scale with influence.

Detection has to move from prose to coordination

Text classifiers become less useful as models generate plausible language cheaply. The stronger signals are account creation patterns, shared infrastructure, timing, cross-platform coordination and provenance.

Social platforms therefore have to detect the network, not just the prose. Account-creation patterns, shared infrastructure, timing and cross-platform coordination become more informative as generated language gets cheaper. Agents lower the cost of running fake personas; they do not make coordinated behavior invisible.

Sources

• New York Times, September 18 autonomous influence-campaign report: https://www.nytimes.com/2026/09/18/technology/iran-china-autonomous-ai-influence-campaigns.html
• Anthropic, September 2026 threat-intelligence report: https://www.anthropic.com/threat-intelligence-report-september-2026