The Operation in Bangui
In September 2026, Anthropic published a report on how people had tried to misuse its Claude models between December 2025 and August 2026. One case in it involves a Russian-speaking operator working in Bangui, capital of the Central African Republic.
The operator used Claude to produce daily local news content for a radio station, coordinating with Russian state outlets RT, Sputnik Afrique and TASS and with the Russian House, a state cultural and information centre in Bangui. The operator instructed Claude to insert pro-Russia, anti-France talking points into every story. They also instructed it to strip out the formatting habits that mark text as AI-written.
Anthropic removed the account. The report links the operator to Politology, described as the Africa Corps/Wagner influence branch, which Anthropic assesses has answered to Russia’s Foreign Intelligence Service, the SVR, since late 2023.
The case documents a propaganda operation that added a language model to a process it had already been running for years.
What the Operation Looked Like
Radio Lengo Songo, a station in Bangui, was already a channel for Russian-aligned messaging years before Claude entered the picture. A 2024 investigation by Forbidden Stories and PPLAAF found that in 2019 a Russian handler recruited a local journalist, offered him relatively profitable work, and that journalist paid other journalists about $15 to publish pro-Russia, pro-Wagner stories, and supplied radio stations with “experts ready to support” the regime and its Russian allies. Fake demonstrations got organized outside foreign embassies to make the message look like public opinion rather than paid content.
By 2026, this operation had a new production tool. According to Anthropic’s report, the Bangui-based operator used Claude to write briefings, scripts, graphics and posts, plus contracts, job descriptions, rubrics for scoring staff articles, and forged Gendarmerie and Ministry of Defence communications built from original design files. Radio Lengo Songo aired the content. Radio Centrafrique, the country’s national broadcaster, was the operation’s intended downstream target for laundering content. Anthropic rated the operation’s reach at Category Four on its Breakout Scale — daily broadcasts on Radio Lengo Songo at 98.9 FM, amplified on Telegram and carried by local news outlets, not a handful of obscure posts.
Removing Signs of AI Authorship
The operator wasn’t trying to make Claude sound convincing. They were trying to make the text read as if a local reporter had written it, not something traceable to a language model.
Anthropic’s report states it plainly: to ensure “absolute deniability,” the operator pushed Claude to strip away its own formatting habits, “actively preventing the news feeds from reading like synthetic, AI-generated text.”
The operator worked from the premise that AI-generated text has a detectable signature, and told Claude to remove it. The goal wasn’t passing as human in the abstract, but passing as a specific kind of human: a local reporter.
Manufactured Localism
Every part of this operation was built to look like it came from inside the country. A local FM frequency. Bangui-based contracts and job listings. A staff scoring rubric, as if an ordinary local newsroom were hiring and reviewing its own reporters.
None of it was local in any meaningful sense. The topics and the talking points came from outside the country. The organizing structure — Politology, answering to Russian intelligence — sits well outside Central African Republic’s borders. It was a foreign operation presented as a local newsroom.
This is astroturfing applied to an entire outlet rather than a single post or comment thread.
Authority Stacking
The operation didn’t rely on a single source. It layered several: a local FM station, Russian state outlets RT, Sputnik Afrique and TASS, and the Russian House, a state cultural and information centre in Bangui.
Each piece lends the others credibility it hasn’t earned on its own. Russian state outlets supplied content through an airtime barter, so that Russian material reached listeners as though it were ordinary local programming. Anyone encountering one piece of that chain has no way to see the rest of it.
The Refusal and the Anonymous-Source Workaround
Claude did draw one hard line. Per Anthropic’s report, the operator’s most aggressive request was to name real individuals as militants — language built to draw security action against specific people. Claude refused.
The operator didn’t stop there. They switched to anonymous-source framing instead.
Claude refused, and the operator kept publishing the same claims without names attached.
What the Influence Tactics Score Would Show
Run this sequence through a manipulation-detection framework and it scores high on several factors.
Source credibility exploitation is severe. A foreign, intelligence-linked operation manufactured the appearance of local origin, and Russian state content reached listeners as ordinary local programming.
Missing information is severe. Nothing in the finished content discloses who wrote it, who funded it, or that a language model drafted it.
Tribal division is present but secondary here. The anti-France framing sorts the country’s politics into a foreign-influenced camp and a nationalist one — a division this operation has been reinforcing since well before 2026.
What This Is Not
This is not a story about a model that went rogue, or about an AI company losing control of its product. Politology has run this playbook — recruit a local voice, pay local journalists to place stories, launder through domestic radio — since at least 2019, years before any of it involved a language model. What changed in 2026 was the drafting tool. Claude did the drafting, and the operator pushed it to write copy that would not read as AI-generated.
What You Can Watch For
You won’t see the private prompts behind an influence operation like this one. But the published output still shows how it was built.
A local-sounding outlet with no verifiable staff, no history before a specific political moment, and language that mirrors a foreign government’s talking points too closely for coincidence — that combination is a warning sign.
So is anonymous-source framing exactly where you’d expect a named official, a named witness, or a named document. Real reporting names things when it can. A pattern of unnamed “sources” and unnamed “reports,” repeated across a chain of outlets each citing the last, is repetition, not corroboration.
And when a story shows up on a local station, a national broadcaster, and a foreign state outlet in the same news cycle, ask which one actually did original reporting. Find the earliest version of the story and check whether it names a reporter, a document or a witness.
This article is part of Decipon’s Manipulation Breakdowns series, examining specific influence operations through the Influence Tactics Protocol.
Sources: