AI-Generated Fake News Network Promoted Awami League Using 29 Claude Accounts
A Bangladesh-based automated influence network used artificial intelligence (AI) company Anthropic’s chatbot Claude to produce at least 300 fabricated reports and 1,500 image-generation prompts in Bangla over a period of around 16 months. The operation was reportedly run by a single individual in Gaibandha, who rotated between 29 accounts to generate fake news using AI, with the aim of promoting the Awami League.
Anthropic disclosed the findings in its report, “Detecting and Countering Misuse of AI: September 2026,” published on Thursday, September 10. The company identified the operation as “GTG-54006.”
According to Anthropic’s investigation, the network primarily targeted Bangladesh’s rural population. Its content was largely created in support of the Awami League and against the party’s political opponents. Since the Awami League has been out of power following the July 2024 mass uprising, Anthropic assessed that the operation was not a pro-government campaign but rather an effort conducted from an opposition political position.
Automated Content Across 29 Accounts
Anthropic said an individual from Gaibandha district used 29 Claude accounts on a rotating basis to evade detection. The operation, which continued for around 16 months, relied on custom software called “Fake_News_3.py.” Directly connected to Claude, the software generated 15 headlines, three detailed fabricated reports and 15 image-generation instructions or prompts in each batch.
Anthropic said it found evidence that at least the third version of the software had been used.
According to the company’s estimates, the operation generated at least 1,500 headlines, 300 fabricated reports and 1,500 image-generation prompts in total. Since Claude does not directly offer image-generation capabilities, Anthropic believes these prompts may have been used with other AI models to create the images.
The generated content was initially stored in a designated cloud-storage folder. It was then converted into audio and video content and prepared for systematic publication on Facebook, YouTube and TikTok.
Political Disinformation Built Around Four Narratives
Anthropic’s analysis identified four major narratives in the network’s content.
First, the BNP, Jamaat-e-Islami and National Citizen Party were portrayed as planning to establish Taliban-style rule and infiltrate the security forces.
Second, fabricated stories about assassination plots and hit squads were circulated against leaders of the then-interim government and the student movement.
Third, student leaders of the July uprising were portrayed as agents of foreign intelligence agencies.
Fourth, fabricated allegations of financial corruption and secret alliances were spread against opposition political parties.
An internal communication cited in Anthropic’s report indicates that those involved were aware they were spreading false information. The content was designed to be aggressive but easy to understand and appear credible to people with limited access to education.
Automated Publication Through YouTube API
Another part of the network’s technical infrastructure was capable of bulk-uploading videos through YouTube’s API. The script was operated through a third-party continuous integration service, allowing videos to be published according to schedules set as far as a month in advance.
According to Anthropic, using the third-party service could also conceal the internet protocol (IP) address of the actual operator.
The company further said some of the network’s narratives were consistent with pro-India geopolitical interests, but it found no evidence of state direction or funding behind the operation.
Attempt to Spread Content Across Three Platforms
Under Anthropic’s “breakout scale” classification, the operation was placed in the third category. This was because videos matching the network’s content were identified across multiple Bangladesh-focused channels and profiles on Facebook, YouTube and TikTok.
However, it was not possible to determine exactly which channels had published the full range of content. Anthropic also said it found no evidence that the content had spread on a broader scale beyond the identified accounts.
All Claude Accounts Shut Down
Anthropic said that after identifying the network through an internal investigation, it shut down all Claude accounts associated with the operation. The company also developed a new detection system based on the network’s behavioural characteristics to identify and prevent attempts to create similar accounts.
In addition, Anthropic shared relevant evidence with social media platforms and other partner organisations. Some technical information, including the identity of the cloud-storage folder and details of the upload script, was withheld from public disclosure for security reasons and shared only with partners.
Anthropic’s September report details a total of nine cases of AI misuse. These include influence operations conducted from Russia, Iran, Turkey, the Gulf region, South Asia, Africa and Europe.
According to the report, the Bangladesh case was unusual among them. Rather than a state institution, a single individual used AI for political purposes to conduct a relatively large-scale content-generation operation, Anthropic said.
The company said detecting and disclosing misuse of Claude is part of its responsibility. Over the past eight months, it has taken action against and banned accounts involved in various forms of misuse, including cyberattacks, disinformation and propaganda, surveillance, model theft and biological misuse.
According to Anthropic, AI is now enabling less-skilled individuals and small groups to conduct cyberattacks and disinformation campaigns on a much larger scale. To address these risks, the company said governments, researchers and other AI companies need to strengthen information sharing and coordinated security measures.
//DBTech/SME/MIS/BRB/OR//
A Bangladesh-based automated influence network used artificial intelligence (AI) company Anthropic’s chatbot Claude to produce at least 300 fabricated reports and 1,500 image-generation prompts in Bangla over a period of around 16 months. The operation was reportedly run by a single individual in Gaibandha, who rotated between 29 accounts to generate fake news using AI, with the aim of promoting the Awami League.
Anthropic disclosed the findings in its report, “Detecting and Countering Misuse of AI: September 2026,” published on Thursday, September 10. The company identified the operation as “GTG-54006.”
According to Anthropic’s investigation, the network primarily targeted Bangladesh’s rural population. Its content was largely created in support of the Awami League and against the party’s political opponents. Since the Awami League has been out of power following the July 2024 mass uprising, Anthropic assessed that the operation was not a pro-government campaign but rather an effort conducted from an opposition political position.
Automated Content Across 29 Accounts
Anthropic said an individual from Gaibandha district used 29 Claude accounts on a rotating basis to evade detection. The operation, which continued for around 16 months, relied on custom software called “Fake_News_3.py.” Directly connected to Claude, the software generated 15 headlines, three detailed fabricated reports and 15 image-generation instructions or prompts in each batch.
Anthropic said it found evidence that at least the third version of the software had been used.
According to the company’s estimates, the operation generated at least 1,500 headlines, 300 fabricated reports and 1,500 image-generation prompts in total. Since Claude does not directly offer image-generation capabilities, Anthropic believes these prompts may have been used with other AI models to create the images.
The generated content was initially stored in a designated cloud-storage folder. It was then converted into audio and video content and prepared for systematic publication on Facebook, YouTube and TikTok.
Political Disinformation Built Around Four Narratives
Anthropic’s analysis identified four major narratives in the network’s content.
First, the BNP, Jamaat-e-Islami and National Citizen Party were portrayed as planning to establish Taliban-style rule and infiltrate the security forces.
Second, fabricated stories about assassination plots and hit squads were circulated against leaders of the then-interim government and the student movement.
Third, student leaders of the July uprising were portrayed as agents of foreign intelligence agencies.
Fourth, fabricated allegations of financial corruption and secret alliances were spread against opposition political parties.
An internal communication cited in Anthropic’s report indicates that those involved were aware they were spreading false information. The content was designed to be aggressive but easy to understand and appear credible to people with limited access to education.
Automated Publication Through YouTube API
Another part of the network’s technical infrastructure was capable of bulk-uploading videos through YouTube’s API. The script was operated through a third-party continuous integration service, allowing videos to be published according to schedules set as far as a month in advance.
According to Anthropic, using the third-party service could also conceal the internet protocol (IP) address of the actual operator.
The company further said some of the network’s narratives were consistent with pro-India geopolitical interests, but it found no evidence of state direction or funding behind the operation.
Attempt to Spread Content Across Three Platforms
Under Anthropic’s “breakout scale” classification, the operation was placed in the third category. This was because videos matching the network’s content were identified across multiple Bangladesh-focused channels and profiles on Facebook, YouTube and TikTok.
However, it was not possible to determine exactly which channels had published the full range of content. Anthropic also said it found no evidence that the content had spread on a broader scale beyond the identified accounts.
All Claude Accounts Shut Down
Anthropic said that after identifying the network through an internal investigation, it shut down all Claude accounts associated with the operation. The company also developed a new detection system based on the network’s behavioural characteristics to identify and prevent attempts to create similar accounts.
In addition, Anthropic shared relevant evidence with social media platforms and other partner organisations. Some technical information, including the identity of the cloud-storage folder and details of the upload script, was withheld from public disclosure for security reasons and shared only with partners.
Anthropic’s September report details a total of nine cases of AI misuse. These include influence operations conducted from Russia, Iran, Turkey, the Gulf region, South Asia, Africa and Europe.
According to the report, the Bangladesh case was unusual among them. Rather than a state institution, a single individual used AI for political purposes to conduct a relatively large-scale content-generation operation, Anthropic said.
The company said detecting and disclosing misuse of Claude is part of its responsibility. Over the past eight months, it has taken action against and banned accounts involved in various forms of misuse, including cyberattacks, disinformation and propaganda, surveillance, model theft and biological misuse.
According to Anthropic, AI is now enabling less-skilled individuals and small groups to conduct cyberattacks and disinformation campaigns on a much larger scale. To address these risks, the company said governments, researchers and other AI companies need to strengthen information sharing and coordinated security measures.
//DBTech/SME/MIS/BRB/OR//
