One Person, a Script, and a Queue
Anthropic’s September 2026 threat report describes one person in Gaibandha district, Bangladesh, who used Claude to run a fake news operation.
The operator did not have a newsroom, a budget line or a state sponsor. According to press coverage of the report, the operator had 29 rotating Claude accounts and custom software.
The output was fabricated news in Bengali, aimed at supporters of the Awami League. This article breaks down what the operation did and which of its tactics you can recognize elsewhere.
One limit first. I have read press coverage of the report, not the report itself. Where outlets disagree, I say so. The two outlets differ on the script’s name, and only one gives a duration.
What Was Found
The outlets give the same counts, with different qualifiers on the headlines. They differ on the script’s name, and only one gives a duration:
- Accounts: 29 Claude accounts, rotated to stay under platform limits and avoid detection.
- Output: 1,500 fake headlines (“at least” per National Herald, “around” per The Financial Express), 300 fabricated narratives and 1,500 image prompts.
- Duration: roughly 16 months, according to The Financial Express, which quotes the report. National Herald does not give a timeframe.
- Method: a script named “fake_news_3.py” that produced fixed batches of 15 headlines, three fabricated stories and 15 image prompts. The Financial Express spells the name “fake_news3.py”; National Herald spells it “fake_news_3.py”.
- Distribution: a second script uploaded videos to YouTube on a schedule set weeks or months in advance. Promotional livestreams ran on Facebook and TikTok.
Anthropic banned the accounts, added detection measures, and shared identifiers with platforms and partners. Anthropic found no evidence that the campaign reached a significantly wider audience beyond the accounts linked to the operation, according to The Financial Express.
1. Fabricated News in News Format
The content was not commentary. It was written to look like reporting: a headline, a short story, an image.
That format carries an assumption. A reader who sees a headline and a thumbnail assumes someone checked. Nobody did. According to the report’s account of internal communications, the operator wrote that “no one knows the news is fake.”
The operator was not confused about what was being produced. The same communications reportedly call the material “fake news.” That removes the defense that the content was sloppy or sincere.
2. Accusations Built to Provoke
The fabricated stories targeted the BNP, Jamaat-e-Islami, the National Citizen Party, the interim government and student leaders. The claims alleged ties to foreign agents, plans for Taliban-style rule, murder plots and corruption.
These are accusations that are hard to disprove and easy to repeat. Each one names a threat and an enemy.
Press coverage quotes the operator saying the stories were made “hot and aggressive.” That is a description of an emotional target. The goal was a reaction, and accuracy was not a requirement.
In Influence Tactics terms this is emotional manipulation layered on fabricated content. The stories relied on fear and anger instead of evidence.
3. Writing for the Audience Least Able to Check
The operator reportedly wrote in simple language “so that even people in villages could understand.” The audience was Awami League supporters in rural areas with limited literacy.
Simple language is not manipulation. Plain writing is good. The problem is the pairing: simple wording, alarming claims, and an audience with few ways to verify.
A reader in a village cannot easily call a newsroom or search an archive. Livestreams and short videos add another barrier. A spoken claim is harder to look up than a written one.
This is a targeting decision. The operator chose an audience with few ways to verify claims, and wrote claims simple enough to accept without checking.
4. Scripted Batch Production and Account Rotation
The script is the core of the case. Fixed batches of headlines, stories and image prompts turn writing into a loop. A second script turns the output into scheduled uploads.
Previously, volume required people. A person writing 300 narratives would need weeks and would tire. A script does not.
The 29 accounts show the same logic applied to enforcement. When one account hits a limit, the next takes over. Platform rules assume one user is one person at human speed. That assumption failed here.
The upload scheduling matters too. Videos queued weeks or months in advance mean the operation can run without the operator present. Activity is not a reliable sign of an active person.
5. Hiding Where It Came From
A third-party coordination service helped mask the operator’s IP address, according to The Financial Express. Hiding origin is a standard part of an influence operation.
For a domestic operation the concealment has a plain purpose. Viewers should believe the content comes from independent news channels, not from one person with a script.
If several “independent” channels post the same story in the same format, look for one operator behind them.
Who Was Behind It
Anthropic found no evidence that the Awami League operated or funded the network. The party was ousted from power after the July 2024 uprising. Anthropic categorized the activity as opposition-linked, not state-sponsored.
That distinction matters. It would be easy to read this as a party campaign. The evidence reported so far shows an individual operator working in support of a political side.
It also shows how small the operation was. This one needed one person, 29 accounts and a few scripts.
What the Influence Tactics Score Would Show
Based on press descriptions of the content, not the content itself, three Influence Tactics Protocol factors would likely score high:
- Emotional Manipulation: high. Content was engineered to be aggressive and alarming.
- Missing Information: very high. The stories had no sources because the events did not happen.
- Source Credibility Exploitation: high. News formatting and the appearance of many independent channels borrowed trust the content had not earned.
The score describes the content. It cannot tell you how many people believed it. Anthropic found no evidence of wide reach, and its findings stop at the accounts it could see.
What This Case Does Not Show
It does not show that AI made the operator’s politics more persuasive. It does not show a large audience was fooled. It shows that one person could produce fabricated news over many months, 16 months by one outlet’s count, at a volume that once needed a team.
Implications
Production cost fell. And in this case the operation was detected by the AI company whose model wrote the content. Anthropic found this operation through its own accounts. The coverage does not say whether the platforms that hosted the videos had flagged it.
That is useful, and it has limits. A provider sees only its own accounts. An operator who switches tools leaves no trace there.
For readers, the practical result is that volume is no longer evidence of an organization. A channel that posts daily with consistent quality may be one person and a script.
What to Check
When a news channel or video page shows up in your feed, run four checks:
- Look for a named person or outlet. Find a real masthead, a physical address or a reporter you can search.
- Search the specific claim. If a murder plot or foreign-agent story appears nowhere else, treat it as unconfirmed.
- Notice whether the story is written to make you angry or afraid. Content written to be “hot” is built for sharing, not for informing.
- Check how the channel started. A page with months of scheduled uploads and no history before them is worth distrusting.
If you help older or less-connected relatives with their phones, show them the second check. It takes one search and works in any language.
This article is part of Decipon’s Manipulation Breakdowns series, examining specific influence operations through the Influence Tactics Protocol.
Sources:
Anthropic’s threat report itself is not linked here; the writer did not read it, and everything above comes from press coverage.
- Anthropic shuts down AI network that churned out pro-Awami League fake news — The Financial Express (Bangladesh)
- From fake news to missiles: Anthropic details misuse of AI model Claude — National Herald India