On September 10, Anthropic, the company behind the popular AI assistant Claude, announced that it had disrupted a disinformation operation powered by its own technology. The operation, described in a report titled “Detecting and countering misuse of AI: September 2026,” was not the work of a state intelligence agency or a sprawling political campaign. It was run by a single operator in the Gaibandha district of Bangladesh, using 29 Claude accounts over roughly 16 months. The operator’s goal was to produce a steady stream of fabricated Bengali news stories supporting the Awami League and attacking its political opponents. Anthropic’s investigators found that the operation generated at least 1,500 headlines, 300 false narratives, and 1,500 prompts for images. The material was designed for distribution through social media and video platforms, especially Facebook Live, YouTube, and TikTok. What makes this case so striking is the combination of scale and simplicity. The operator did not need a large team or a big budget. This person used Claude as a kind of automated ghostwriter, feeding it instructions and receiving ready-made propaganda in return. The content was not subtle. Anthropic said the operator knowingly presented fabricated material as genuine news. In communications examined by the company, the operator wrote that people would not know that the news was fake. That single sentence captures the brazenness of the operation. It also highlights a broader truth about AI in the wrong hands: the technology can be used to manufacture lies that look like journalism, and to do so at a pace no human writer could match. In the past, a campaign like this would have required a newsroom of writers, editors, and translators. Now, a single person with an internet connection and a handful of accounts can do it alone. The report does not reveal the operator’s identity, but it paints a picture of someone who was determined, organized, and careful. This person varied prompts, rotated accounts, and tested the limits of the system. The operator was not a hacker breaking into a network; the operator was a user of a commercial product, exploiting it for a purpose its creators never intended. That is what makes AI safety so difficult. The same models that are trained to be helpful and harmless can be redirected by a determined user to produce harmful content. Anthropic’s detection systems eventually caught the activity, but only after months of use. The case is a reminder that AI companies are in a constant race against misuse, and that the race is not always won quickly.
The political content of the operation was clear and aggressive. Anthropic’s report describes a network of fabricated stories that were uniformly pro-Awami League and sharply hostile to a range of opponents: the Bangladesh Nationalist Party, Jamaat-e-Islami, the National Citizens Committee, the interim government, and student protest leaders. The fake articles and headlines accused these groups of being foreign agents and of advancing Taliban-style governance. These are not mild criticisms. In a country with deep political divisions and a history of religious and nationalist tensions, such accusations can be dangerous. The operator was not trying to persuade through reasoned argument; the operator was trying to inflame. Anthropic noted that the instructions embedded in the operation called for material that was “hot and aggressive” and written in simple language that would be easily understood by rural audiences. This is a telling detail. It shows that the operation was aimed at ordinary people, not political elites. The use of Bengali, the focus on short videos, and the emphasis on emotional language all point to a strategy of reaching voters where they are and speaking to them in a way that feels familiar and trustworthy. The goal was not to inform but to provoke. The operator wanted to make people angry at political rivals, suspicious of the interim government, and loyal to the Awami League. At the same time, Anthropic found no evidence that the Awami League directed or financed the operation. Some narratives, the company said, were consistent with pro-Indian geopolitical interests, but investigators found no evidence that any government was behind it. In other words, this was not a state-sponsored conspiracy. It was the work of an individual with a strong political agenda and a willingness to use AI to amplify it. That makes the case both more ordinary and more disturbing. It is easy to imagine similar operations being run by people in other countries, with other languages, other parties, and other grievances. The specific politics may change, but the pattern is the same: use AI to create a flood of content that is designed to divide, anger, and manipulate.
Behind the political drama lies a technical story that reveals how AI can be weaponized with surprising ease. Anthropic’s report describes a programme called “fake_news_3.py” that was configured to connect directly to Claude and request a fixed batch of material. Each request produced 15 Bengali headlines, three detailed fabricated stories, and 15 English-language prompts for accompanying images. This was not a one-off experiment. The programme was designed to run repeatedly, generating a continuous supply of propaganda. Once the text and image prompts were ready, they were passed through cloud-storage folders and converted into audio and video. A separate programme automated bulk uploads to YouTube and scheduled their publication, so the content would appear at chosen times, as if a real editorial team were behind it. The operator also used a third-party software automation service that concealed the operator’s internet address, making it harder for platforms to trace the activity back to its source. Anthropic noted that the image prompts were probably sent to other AI models, since Claude itself does not generate images. The company did not identify which other AI services may have been used. This detail matters because it shows that the operation was not locked into a single tool. It combined different AI systems, cloud storage, automation software, and social media platforms into an assembly line of deception. The whole setup was designed to evade detection: multiple accounts, hidden IP addresses, automated scheduling, and a relentless production cycle. For someone unfamiliar with AI, this may sound like science fiction. But it is happening now, with tools that are freely available online. The fact that one person in a small district in Bangladesh could build such a pipeline, without apparent state support, is a sobering reminder of how accessible these capabilities have become. It also shows that the threat is not limited to deepfakes or sophisticated hacking. Sometimes the most effective disinformation is simple text, written in a familiar language, packaged as video, and distributed through ordinary platforms. The automation made it possible to produce content faster than any human team could, and the use of multiple accounts made it harder for platforms to connect the dots. In the end, the operation was detected, but only after a long period of activity. That suggests that similar operations may be running right now, unnoticed, in other parts of the world.
Anthropic’s response was swift, but it also reveals the limits of what any single company can do. After an internal investigation, Anthropic banned the 29 Claude accounts associated with the activity. It shared technical information with distribution platforms and other partners, presumably including YouTube, Facebook, and TikTok, so they could take their own action. It also introduced additional detection measures intended to prevent the operator from simply opening new accounts and starting again. These are sensible steps, and Anthropic deserves credit for being transparent about the case. But the report does not establish how widely the fabricated material circulated, how many people viewed it, or whether it influenced political behavior. That is a crucial gap. We do not know if the fake news reached thousands of people or millions. We do not know if it changed anyone’s mind, reinforced existing beliefs, or contributed to real-world tensions. In the absence of such data, it is hard to measure the actual harm. What we do know is that the operation was active for about 16 months. That is a long time for false narratives to sit on social media, accumulate views, and be shared by unsuspecting users. Even if the content was not widely seen, the attempt itself is significant. It shows that AI can be used to manufacture political propaganda cheaply and quickly, and that the people who do this are learning how to evade safeguards. The response from Anthropic and its partners may have shut down this particular operation, but the same playbook could be used again, with different tools, different languages, and different targets. There is also a question of accountability. The operator may still be active, using other tools or other accounts. Anthropic can ban accounts, but it cannot arrest anyone. Law enforcement and regulators have not yet caught up with the reality of AI-generated disinformation. The company’s report is best understood not as a success story, but as an early warning. It tells us what is possible, and it asks us to think about what should be done before the next operation appears.
To humanize this story, it helps to step back and think about what it means for real people in Bangladesh. The country has been through a period of intense political upheaval, with student-led protests, changes in government, and deep divisions between parties. In such an environment, false information is not just an abstract problem. It can inflame tensions, spread fear, and undermine trust in democratic processes. The fabricated stories described in Anthropic’s report were aimed at rural audiences, people who may not have the time or resources to fact-check every video that appears in their feed. The operator wanted to make them angry, suspicious, and loyal to one side. The phrase “hot and aggressive” is chilling because it reveals an intent to provoke emotional reactions, not to inform. The people who would have watched these videos are not statistics; they are voters, neighbors, parents, and young people trying to make sense of a chaotic political landscape. They deserve better than to be manipulated by an anonymous figure with a chatbot and a grudge. This is the human cost of AI-driven disinformation. It is not just about algorithms and accounts; it is about the erosion of shared reality. When people cannot agree on basic facts, when they are told that their political opponents are foreign agents or religious extremists, the possibility of peaceful dialogue shrinks. The operator in Gaibandha may have acted alone, but the operation was aimed at a collective: the public. And the public, in this case, includes millions of Bengali-speaking people who rely on social media for news. Anthropic’s report does not tell us whether the operation succeeded in shaping opinions. But it tells us that someone believed it could, and that belief alone is enough to cause concern. It also reminds us that disinformation is not a victimless crime. It targets the trust that holds communities together, and it exploits the vulnerabilities of people who are already anxious, uncertain, or afraid.
Ultimately, this episode is a reminder that the rise of AI is not only a story about innovation and convenience. It is also a story about vulnerability. The same technology that can help doctors write notes, students learn, and businesses serve customers can also be used to create lies that look like news. Anthropic’s report is a valuable piece of transparency, but it is only one piece. Governments, tech companies, civil society, and ordinary users all have a role to play in protecting the information environment. Platforms need to invest in detection and work with AI developers to share threat intelligence. AI companies need to monitor how their models are being used and act quickly when they see abuse. Journalists and fact-checkers need to help audiences understand what is real and what is not. And individuals need to be skeptical, especially when content is designed to provoke anger or fear. None of this is easy. Disinformation is a cat-and-mouse game, and the mice are getting smarter. But the alternative is to accept a world where a single person can flood a country with fake news, hide behind automation, and never face consequences. That is not a world we should want. The story of the Gaibandha operation is a warning, but it is also an opportunity. It shows that AI misuse can be detected, that companies can cooperate, and that the public can be informed. Whether that is enough will depend on how seriously we take the threat. For now, the best response is to stay alert, ask questions, and remember that behind every headline, real or fake, there are real people whose lives and communities are affected. In the end, the fight against AI-powered disinformation is not just a technical challenge. It is a human one.