The AI company Anthropic says Russian operatives secretly used its Claude chatbot to generate propaganda for a Russian-funded radio station in the Central African Republic, portraying the country's government and Russian mercenaries in a glowing light while attacking France and the political opposition. The case, described in Anthropic's new report on AI misuse, is one of several pro-Russian influence operations uncovered across Africa — from the Democratic Republic of Congo to Kenya, Sudan and Mali — and shows how quickly generative AI has become a workhorse of foreign disinformation, Deutsche Welle reported.
Table of Contents
- The Bangui Operation
- When the AI Said No
- A Continent-Wide Pattern
- Why It Matters
- How Anthropic Responded
- Key Takeaways
- Frequently Asked Questions
- Sources
- Read Next on Chronicle
The Bangui Operation
The report, titled "Detecting and Countering Misuse of AI" and published in September, details how Claude was used to produce content for Radio Lengo Songo, a station in Bangui that the investigative collective All Eyes on Wagner identifies as Russian-funded, according to DW. The broadcasts painted President Faustin-Archange Touadéra's government and the Africa Corps — the successor force to the Wagner mercenary group — in a positive light, while criticizing France, the former colonial power, and opposition politicians.
Lou Osborn of All Eyes on Wagner, who contributed to the research, told DW the techniques were familiar: "There have been several arrests of people who were not necessarily closely affiliated with the Touadéra regime. They've been doing this for several years now." The Central African Republic has been one of Russia's closest African partners since 2016, when Touadéra took power, hosting Wagner-linked operations that span mining, security and media.
What is new is the industrialization of the messaging. Rather than a team of propagandists writing each script by hand, the operation could generate high volumes of polished content on demand — a force multiplier for influence campaigns in countries with small media markets and limited fact-checking infrastructure.
When the AI Said No
The report contains one striking detail: the operators did not always get what they asked for. When prompted to portray opposition figures as armed militants, Claude refused the request. It is a small but telling data point — the model's safety training occasionally tripped even while the same system was being used to churn out the broader propaganda pipeline. It also shows how influence operators adapt: they worked around the refusal rather than abandoning the tool.
The incident underscores a tension that AI labs are now confronting head-on. A model safe enough to decline one malicious request can still be weaponized in volume for thousands of others, especially when the harmful use is not obviously illegal at the level of a single prompt. Polishing a favorable news script about a government does not look like wrongdoing in isolation; it is the orchestration — fake accounts, fake news sites, coordinated amplification — that turns it into an operation.
A Continent-Wide Pattern
The Central African Republic was not an isolated case. Anthropic says it found similar AI-assisted influence networks in four other countries:
- Democratic Republic of Congo: a network of around 70 fake news websites published 318 AI-generated articles, amplified by more than 250 fictitious accounts on X.
- Kenya: Claude was allegedly used to generate batches of 50 messages supporting the government and attacking the opposition.
- Sudan: AI-generated fabricated witness statements were created for the United Nations.
- Mali: an adviser with intelligence-service ties used Claude to build a platform for monitoring telecommunications.
Together, they sketch a playbook: synthetic content at scale, distributed through disposable websites and fake personas, aimed at societies already strained by conflict or political competition. DW reported that Anthropic discovered and shut down the operations between December 2025 and August 2026 by deleting the accounts involved and strengthening its detection systems.
Why It Matters
The findings arrive at a moment when governments on both sides of the Atlantic are scrambling to define what "AI safety" means in practice. In the United States, the Trump administration has just stood up a new AI task force and the FTC has opened its first investigation into rogue AI agents — developments chronicled in our recent coverage of the administration's "Super Intelligence Force".
The Anthropic report shifts part of the debate from what AI might do in the future to what it is doing right now: enabling state-backed actors to produce convincing, locally tailored propaganda at a fraction of the historical cost. Analysts warn that the African cases are a preview rather than an exception. The same tooling can target European or American elections — and, indeed, Moscow's earlier interference efforts suggest those audiences are the ultimate prize.
There is also a strategic irony. Russia's disinformation apparatus, built up over a decade around human troll farms, is being upgraded with American AI. The question for labs like Anthropic is whether detection and account removal — a fundamentally reactive approach — can keep pace with adversaries who treat each ban as a cost of doing business.
How Anthropic Responded
Anthropic says it deleted the accounts involved and hardened its detection systems. The company framed the September report as part of a broader transparency push: naming and describing misuse operations publicly, rather than quietly banning accounts, so that platforms, journalists and governments can recognize the patterns elsewhere.
Critics will note the limits. Shutting down accounts after months of operation still leaves the content circulating — the 318 articles in the DRC, for instance, were already published and amplified. And detection depends on the adversary reusing the same accounts, infrastructure and behavioral tells. More sophisticated actors will diversify across providers and open-source models that have no corporate owner to ban them at all.
For now, the report serves as the most detailed public accounting yet of how a frontier AI model has been folded into a state influence machine — and of the cat-and-mouse game that is only beginning.
Key Takeaways
- Anthropic's September report says Russian operatives used Claude to produce propaganda for Radio Lengo Songo in the Central African Republic, boosting the Touadéra government and Africa Corps while attacking France and the opposition.
- When asked to portray opposition figures as armed militants, the model refused — showing safety guardrails catching some requests while the broader campaign continued.
- Similar AI-driven operations were found in the DRC (70 fake sites, 318 articles), Kenya, Sudan and Mali; Anthropic shut the accounts between December 2025 and August 2026.
- The case shows generative AI industrializing foreign influence at low cost, raising urgent questions about detection, deterrence and the limits of corporate self-policing.
Frequently Asked Questions
What did Anthropic's report find?
Its report "Detecting and Countering Misuse of AI" says Russian operatives used the Claude chatbot to generate pro-Kremlin content for a Russian-funded radio station in the Central African Republic, and identified similar AI-assisted disinformation networks in the DRC, Kenya, Sudan and Mali.
What is Radio Lengo Songo?
A radio station in Bangui, the Central African Republic's capital, that the investigative collective All Eyes on Wagner identifies as funded by Russia. Anthropic says Claude was used primarily to produce content for its broadcasts.
How did Anthropic respond?
It deleted the accounts involved between December 2025 and August 2026 and strengthened its detection systems, while publishing the report to help others recognize the patterns.
Why does AI make disinformation harder to fight?
Generative AI lets operators produce large volumes of polished, locally tailored content cheaply, distributed through fake news sites and fictitious social accounts — overwhelming the fact-checking capacity of smaller media markets.
Sources
- Deutsche Welle — October 2026: "Anthropic report: Is Russia using AI for disinformation in the Central African Republic and elsewhere?"
- AI Startups News — October 4, 2026: "Anthropic finds Russia used Claude AI to spread propaganda in Central African Republic."
- Europe Says — October 4, 2026: "Anthropic: Is Russia using AI for disinformation in CAR?"
