Anthropic Says It Removed Claude Accounts Tied to Nine Influence Operations

Anthropic says it removed accounts tied to nine influence operations that used Claude to create deceptive political material and the systems needed to distribute it. The cases stretched across six continents, showing how an AI assistant can support both fake content and the organizational work behind it.

In its September threat report, Anthropic said the operations originated in Russia, Iran, Turkey, the Gulf, South Asia, Africa, and Europe. It attributed the activity to a mix of governments, state-aligned propaganda institutions and media, private influence firms, domestic political operators, and an exiled opposition movement. Several campaigns were timed around elections, including fabricated claims about Moldova’s president before its September 2025 vote and fake grassroots posts prepared ahead of Kenya’s 2027 election.

From posts to an operating system

The reported activity went beyond prompting a model for a single post. Anthropic said actors built networks of fake social-media profiles and entire news sites, then used deceptive personas and impersonations of real people and institutions to hide who was behind the material. The company also found actors using persistent files, shared source lists, evasion rules, custom software, and batch calls rather than relying only on isolated prompts.

The reported uses of Claude included

• Creating contracts, job descriptions, staff-scoring rubrics, scripts, graphics, posts, and political talking points for a Central African Republic operation.
• Producing forged government documents and building personas, fake news sites, and deceptive political content.
• Turning standing instructions and approved sources into repeatable content workflows.

One campaign reached beyond the platform

Anthropic described its most detailed case as a Russian state-aligned operation in the Central African Republic. It said a Russian-speaking actor in Bangui supplied content for an operation distributed through Radio Lengo Songo and other outlets. Anthropic assessed the campaign as Breakout Scale Category Four, saying its material was broadcast daily, amplified on Telegram, and carried by local news outlets.

That case illustrates the difference between producing content and giving it reach. Anthropic said most content it found drew little or no authentic engagement, and that state-media distribution created the widest authentic reach among the reported operations. In the Central African Republic case, the company said Claude refused a request to name real people as militants in an effort to prompt security action. The actor instead shifted to anonymous-source framing.

Detection before publication, with a boundary

Anthropic’s proposed advantage is timing. The company said planning and content-generation activity can leave detectable signals, allowing it to identify some operations before content reaches social-media platforms. Once it identifies an operation, Anthropic says it removes involved accounts, attributes the activity to the organization behind it, and feeds the tactics it found into automated detection systems.

But the report also sets out a practical constraint: a model provider’s direct view is at the production stage, not across every channel where a campaign may later travel. That makes account removals and behavioral detection useful interventions, while leaving the harder question of how to stop material that has already entered a broader media network.