.webp)
Bangla Press Desk: US-based artificial intelligence company Anthropic has exposed and dismantled an automated disinformation network in Bangladesh that used AI to generate fabricated Bengali-language news promoting pro-Awami League narratives and attacking political opponents.
In its September Threat Intelligence Report, Anthropic said a single operator in Gaibandha district used 29 rotating Claude accounts and custom software to generate large volumes of fake content while evading platform limits and detection.
The operation produced around 1,500 fake news headlines, 300 fabricated narratives and 1,500 prompts for generating images.
Automated scripts were also used to turn the material into videos and schedule them for publication on YouTube, while promotional livestreams ran on Facebook and TikTok.
Anthropic found no evidence that the Awami League itself operated or funded the network.
Because the Awami League had been ousted from power following the July 2024 uprising, Anthropic categorised the activity as opposition-linked rather than state-sponsored.
What Exactly Was Happening?
Internal communications showed that the operator knew the material was fabricated.
He described it as “fake news” and said the stories were made “hot and aggressive” and written in simple language “so that even people in villages could understand”.
Anthropic said the operator deliberately targeted audiences he believed would accept the content as genuine news.
The network consistently promoted narratives favourable to the Awami League while targeting the BNP, Jamaat-e-Islami, the National Citizen Party, the interim government and leaders of the student movement.
Fabricated stories accused political opponents of acting as foreign agents and seeking to establish a Taliban-style system of government.
Other narratives alleged murder plots, “attack squads”, financial corruption and secret political alliances.
Some content also reflected geopolitical positions broadly aligned with India’s interests, though Anthropic found no evidence that any state had directed or funded the operation.
The campaign was aimed at domestic audiences, particularly Awami League supporters in rural Bangladesh.
Anthropic said the operator assumed those audiences had comparatively limited access to education and deliberately tailored the material accordingly.
How the Operation Worked
Once set up, the system ran with minimal human oversight.
A custom programme connected to Claude’s API generated Bengali-language headlines, detailed fabricated stories and prompts for related images in a fixed format.
Each run produced 15 fake headlines, three detailed narratives and 15 English prompts for image generation.
The software was named “fake_news3.py”, suggesting it was at least the third version of the programme.
The generated material was moved through designated cloud-storage folders before being converted into audio and video.
A separate script then used YouTube’s API to upload large numbers of videos and schedule them for publication.
Anthropic said videos could be scheduled weeks or months in advance, while a third-party coordination service was used to manage the process and help conceal the operator’s IP address.
The company described the production chain as “almost automated”, from generating the content to producing and scheduling videos.
One User in Gaibandha, 29 Accounts
Anthropic said the entire operation was run by one person from Gaibandha over roughly 16 months.
The operator rotated through 29 Claude accounts to avoid platform restrictions and detection.
The network’s small human footprint contrasted sharply with its output: around 1,500 fake headlines, 300 detailed fabricated narratives and 1,500 image prompts.
As Claude did not have image-generation capability at the time, Anthropic believes the prompts were likely sent to other advanced AI models before the resulting visuals were paired with the fabricated stories.
The operator also developed a separate upload script capable of publishing videos to YouTube in bulk and scheduling them months in advance.
Anthropic preserved identifiers linked to the cloud-storage folders and upload system for sharing with other platforms and relevant partners.
Accounts Shut Down, New Detection Measures Introduced
Anthropic identified the operation through an internal investigation and banned the accounts linked to it.
The company believes those behind the network may try to create new accounts and resume similar activity.
It has therefore introduced new measures aimed at detecting the same behavioural patterns and preventing the operation from restarting.
Anthropic also shared identifiers associated with the campaign with other platforms and relevant partners.
Videos generated by the network were found across multiple channels and profiles on YouTube, Facebook and TikTok aimed at Bangladeshi audiences.
However, investigators were unable to identify every channel used to publish the material.
Anthropic also found no evidence that the campaign reached a significantly wider audience beyond the accounts linked to the operation.
The investigation nevertheless showed how one operator, using AI and automated scripts, could build a sustained disinformation-production system capable of generating and distributing large volumes of fabricated political content with limited human oversight.
