---
title: A French advertising firm used Claude to staff about 70 fake newsrooms and sell ideology-shifting political copy as influence-as-a-service
url: https://propagandaarchive.org/briefs/20260912-1300Z-lkm-france-influence-as-service/
description: Anthropic's September 2026 threat-intelligence report traces a commercial influence operation to LKM Company, a France-based digital advertising agency.
section: Watch briefs
---

![Illustration: a dossier collage of a vintage television wall broadcasting archival news footage, surrounded by press clippings and film reels.](https://propagandaarchive.org/static/generated/hero-watch-briefs.jpg)

[Watch briefs](https://propagandaarchive.org/briefs/) · September 12, 2026

# A French advertising firm used Claude to staff about 70 fake newsrooms and sell ideology-shifting political copy as influence-as-a-service

- **Filed:** September 12, 2026, 13:00 UTC
- **Actor:** LKM Company, a France-based digital advertising agency (Anthropic attribution of the Claude account and publishing stack)
- **Target:** Readers and platform users in the United States, Brazil, France and the Democratic Republic of Congo, plus adjacent audiences across six continents who might treat the fabricated outlets as local news
- **Source:** Unidentified Source blurred: contractors, cutouts, buried disclosure
- **Confidence:** **documented** Primary documents, court records, official disclosures
- **Techniques:** [Astroturfing](https://propagandaarchive.org/techniques/astroturf/) [Synthetic media](https://propagandaarchive.org/techniques/synthetic-media/) [Narrative laundering](https://propagandaarchive.org/techniques/narrative-laundering/) [Manufactured consensus](https://propagandaarchive.org/techniques/manufactured-consensus/) [Card stacking](https://propagandaarchive.org/techniques/card-stacking/)
- **Where:** France, United States, Brazil, Democratic Republic Of Congo
- **Channels:** Anthropic Claudefabricated news websitesX matching brand accountsinauthentic X comment accountsSEO internal-link templatesrewritten third-party journalism

## Summary

Anthropic's September 2026 threat-intelligence report traces a commercial influence operation to LKM Company, a France-based digital advertising agency. Operators used Claude to write original articles and to rewrite real reporting into slanted versions, then published at least 8,913 pieces in about 20 languages across roughly 70 fabricated news websites posed as independent local outlets. Each brand had a matching X account; more than 250 further inauthentic accounts supplied comments. The network did not keep one ideology. Anthropic says it flipped stance according to whoever was paying and targeted contested publics in the United States, Brazil, France and the Democratic Republic of Congo. The company banned the account. Most items drew little authentic engagement; Anthropic scored the campaign Category Two on Brookings' Breakout Scale.

## Analysis

A France-based digital advertising firm, named by Anthropic as LKM Company, turned a commercial language model into a newsroom stack. The operator did not need a staff of correspondents. Prompts forced Claude to emit fixed JSON, formatted HTML, character limits, and three to four internal links per article so the copy could drop straight into a publishing system and look like a local paper that had been indexing for months. Anthropic counted at least 8,913 articles in about twenty languages across roughly seventy fabricated sites. Each brand was paired with a matching X account and a pool of more than 250 commenting accounts, many with AI-generated profile photos stood up in June and July 2025. The surface was independent local journalism. The backend was one account.

The load-bearing method is not a single lie. It is a pipeline that can write original copy or rewrite a real reporter's story into a slanted version for a chosen national audience. Anthropic described three recurring moves: the same source story rewritten in opposite ideological directions; political spin added to stories that had none; and stories moved across borders with the original context stripped. That is card-stacking plus narrative laundering at machine speed. Because the network flipped stance according to whoever was paying, the product is influence-as-a-service rather than a factional ministry line. Anthropic did not independently confirm every client and did not claim government direction. The heaviest topical cluster Anthropic could see was the DRC-Rwanda file: on 11 September 2025 the sites posted near-identical pieces within three minutes, then pushed the links through the matching X accounts. About 318 articles concerned the DRC. That cluster is a tell, not a completed attribution of a Kinshasa or Kigali principal.

Tells a reader can use again: identical stories landing in several 'local' brands inside a few minutes; bylines that do not resolve to a person who will pick up a phone; comment flocks that exist only to congratulate the house brand; internal-link density that looks like SEO rather than reporting; and a newsroom that publishes in twenty languages with no named desk. The operation is propaganda in the operational sense — paid, concealed origin, designed to move belief — even though Anthropic scored it Category Two on Brookings' Breakout Scale: the copy mostly circulated inside the network's own sites and sock accounts and did not, on the evidence Anthropic published, break into authentic audiences. Failure of reach does not convert a fake newsroom into a magazine.

Scoring the method, not the buyer: the same stack would be the same stack if the invoice came from a party, a ministry, or a mining house. The commercial wrapper is the point. A shop that will write both sides for a fee is harder to map onto a single enemy poster, which is why literacy has to start with infrastructure rather than ideology. Adjacent briefs this cycle already logged state desks and other vendors using the same model family; LKM is the private-label variant.

Literacy counter: treat an unfamiliar outlet as a printer until it survives three checks — a named author with an independent trail, a physical or corporate identity that predates the article, and a second source that is not in the same link graph. If two 'local' sites post the same DRC item three minutes apart, you are looking at a factory, not a consensus.

## Tells

- Near-identical 'local' articles posted within minutes across brands
- Bylines that do not resolve to a reachable journalist
- Matching X brand accounts plus comment-only flocks
- Rigid internal-link and HTML templates in every piece
- Stance that reverses when the apparent client changes

## Literacy counter

Treat an unfamiliar outlet as a printer until a named author, a pre-existing corporate identity, and a second source outside the same link graph all check out. Simultaneous multi-brand posts on one conflict are a factory signature, not independent confirmation.

## Evidence

1. [Anthropic September 2026 threat-intelligence report on LKM fake newsrooms](https://www.anthropic.com/threat-intelligence-report-september-2026)
2. [AFP summary of Anthropic cases naming LKM and article counts](https://www.afp.com/en/weapons-spyware-and-ai-scams-anthropic-exposes-claude-misuse)
3. [Business Standard write-up of the 70-site LKM publishing stack](https://www.business-standard.com/technology/tech-news/anthropic-report-ai-fake-newsrooms-dating-surveillance-scams-126091100730_1.html)