Countering AI Misuse: September 2026

In September 2026, Anthropic disclosed cases of state-linked propaganda campaigns that exploited its AI models and suspended the associated accounts. These measures reflect the results of detecting and responding to attempts to use AI to produce large-scale disinformation with small teams or to achieve specific political objectives.

State Media Pipelines and the Moldova Case

Evidence was found that Sputnik Moldova, Sputnik en Español, RT, and others used Claude to produce large-scale propaganda, including pre-election candidate smearing, subtitle generation, and voiceovers. In response, Anthropic suspended four related accounts and shared detection indicators.

Iran-Linked Campaign (GTG-34001)

State-linked entities in Iran, including ICCO, the Khorasan Razavi Propaganda Department, and the Bina Observatory, were found to have planned and executed 'soft war' campaigns using Claude. Based on Jihad al-Tabyin, they generated content in 20 languages and attempted persona spoofing and source laundering. Notably, they impersonated sources from reputable institutions such as CSIS and distributed content through the Manjanegh factory to over 100 IRGC-linked channels including Eitaa and Telegram. Although they attempted to evade detection using VPNs and foreign phone numbers, their identities were confirmed through analysis of conversation logs and document footers, leading to account suspensions. This was assessed as Breakout Scale Category 3.

Bangladesh Disinformation Campaign (GTG-54006)

In Bangladesh, Claude was exploited to generate fake news in support of the Awami League. An individual in Gaibandha used 29 Claude accounts and a script (fake_news_3.py) to generate 1,500 headlines, 300 false narratives, and 1,500 image prompts. This content was used in Facebook, YouTube, and TikTok live streams to smear opposition figures. While there was no evidence of direct party involvement or state funding, the accounts were suspended due to organized disinformation dissemination targeting rural areas. This was also classified as Breakout Scale Category Three.