- Filed
- September 11, 2026, 16:00 UTC
- Actor
- LKM Company (France-based digital advertising agency operating a commercial influence-as-a-service network)
- Target
- Audiences in contested democratic spaces including the United States, Brazil, France, and the Democratic Republic of Congo, plus broader publics across six continents reachable via search and social amplification of the fabricated outlets
- Source
- Unidentified Source blurred: contractors, cutouts, buried disclosure
- Confidence
- high Multiple independent open sources, low disagreement about the facts
- Techniques
- Narrative launderingAstroturfingSynthetic mediaManufactured consensusCard stacking
- Where
- France, United States, Brazil, Democratic Republic Of Congo, Global
- Channels
- Anthropic Claude (article generation, rewriting of real journalism into slanted versions, fixed JSON/HTML output pipelines)
- Approximately 70 fabricated news websites on shared infrastructure registered from France
- 70 linked matching X/Twitter accounts
- Network of more than 250 inauthentic commenting X accounts with AI-generated profile photos
- Search-engine optimization and internal linking designed to boost site authority
Summary
Anthropic’s 10 September 2026 threat-intelligence report (GTG-54002) documents a commercial ‘influence-as-a-service’ operation traced to LKM Company, a France-based digital advertising agency. The actor used Claude to mass-produce and rewrite political content for approximately 70 fabricated news websites designed to appear as independent local newsrooms, paired with 70 matching X accounts and more than 250 inauthentic commenting accounts that used AI-generated profile photos. At least 8,913 articles appeared in about 20 languages; the network shifted ideological stance according to paying clients and focused early output on the DRC-Rwanda conflict in support of Kinshasa’s position on mineral deals. Anthropic found no evidence of government direction, disrupted the activity before authentic audience breakout (Brookings Breakout Scale Category Two), and banned the account.
Analysis
Primary source is Anthropic’s ‘Detecting and Countering Misuse of AI: September 2026’ report, case GTG-54002, published 10 September 2026 and covering activity disrupted between December 2025 and August 2026. Anthropic’s Threat Intelligence team identified an account that used Claude to sustain a commercial publishing and amplification system whose public surface consisted of roughly 70 websites styled as independent local newsrooms. Each site was paired with a matching X account; a further layer of more than 250 inauthentic commenting accounts, many created in June–July 2025 and using AI-generated profile photos, supplied engagement. Claude was used both to write original articles and to rewrite genuine reporting into ideologically slanted versions tailored to specific national audiences and political angles. Output was constrained to fixed JSON structures, formatted HTML, character limits and internal-link quotas so that the model functioned as an automated production line rather than a conversational tool. Bylines carried the names of non-existent journalists, giving each outlet the appearance of a staffed local newsroom.
Mechanism: narrative-laundering is load-bearing. Real or invented stories are moved through the institutional form of ‘local news’ so that origin (a commercial French advertising firm selling influence) is invisible to the ordinary reader or search user. Astroturf supplies the social proof via the commenting layer and the coordinated simultaneous publication of near-identical DRC-Rwanda pieces across multiple sites within minutes on 11 September 2025. Synthetic-media enables both the volume of 8,913+ articles in twenty languages and the AI-generated profile imagery. Manufactured-consensus is attempted by density of converging outlets that can be treated by third parties or retrieval systems as independent sources. Card-stacking appears in the selective rewriting of source material—adding political angles where none existed, reversing ideological direction for different paying audiences, and stripping original context when stories are moved across borders.
This qualifies as propaganda and an influence operation because institutional form is performed, commercial sponsorship is concealed from the target audiences, and the design goal is durable shaping of information environments for paying clients rather than transparent advocacy under an LKM byline. Anthropic explicitly frames the activity as a commercial influence-as-a-service model that shifts stance according to whoever is paying; no evidence of state direction was found. Technique scoring remains actor-agnostic: identical fabricated-newsroom factories, AI rewriting pipelines, and inauthentic amplification layers appear in both state and commercial campaigns already documented in the same Anthropic report and in prior FRAME briefs. Color is gray: ultimate commercial provenance is recoverable from Anthropic’s investigation, yet the content surfaces that reach audiences omit that provenance. The operation was disrupted early and remained at Breakout Scale Category Two, generating little observable authentic engagement.
Literacy counter: when a cluster of previously unknown ‘local’ news sites suddenly appears with polished multi-language coverage of contested geopolitical or electoral topics, shares infrastructure, publishes near-identical stories within minutes, and is amplified by accounts using AI-generated profile photos, inspect domain registration patterns, WHOIS, hosting, staff pages, physical addresses, and any commercial advertising-agency linkages. Prefer named journalists with verifiable histories, primary documents, and independent local reporting. Demand that provenance tags travel with every extracted paragraph or social post. Apply the identical fabricated-outlet and commercial influence-as-a-service test to every faction’s AI-augmented publishing pipelines—the method does not change with the client.
Tells
- Cluster of previously unknown ‘local’ news sites publishing near-identical stories within minutes on contested topics
- Shared hosting and domain-registration patterns concentrated in a short window
- Bylines naming journalists who do not exist
- Amplification layer of accounts created in the same timeframe using AI-generated profile photos
- Ideological stance that shifts according to client rather than fixed doctrine