AI chatbot alerted FBI after two Pakistani teens used it to plan murders, report says

Two 17-year-old boys in Pakistan consulted Anthropic’s AI chatbot Claude while planning murders modelled on American crime dramas, and Anthropic reported their activity to the Federal Bureau of Investigation, according to a report by The News published by Geo News.

The two boys — one in Bahawalpur and one in Karachi, 820 kilometres apart — do not know each other, but both are college students said to be heavily influenced by the dramas Breaking Bad and Dexter, the report said. Anthropic conveyed its findings to the FBI, which passed the information on to Pakistani authorities. The Karachi case is being revealed for the first time by The News.

The Bahawalpur boy wanted to kill his father after a disagreement, the report said. He developed an interest in chemistry after watching Breaking Bad and set up a small home lab, obtaining chemicals from Australia and Islamabad, then attempted to poison his father by mixing laboratory-grade methanol into his food. When the attempt failed, he consulted Claude’s chatbot on 12 September, telling it his father had “accidentally” ingested 15–20ml of methanol mixed with food but that nothing happened, and asked the chatbot to explain why. He also claimed to know how to make a dangerous poison.

When Claude refused to guide him further, the boy reportedly wrote: “F*** off. I have Grok.” He is reported to have then consulted xAI’s Grok and OpenAI’s ChatGPT as well, but only Claude alerted the FBI, the report said. A court has since granted him bail after his father forgave him, even if he had been planning something sinister.

The Karachi boy, who was said to have modelled himself on Dexter Morgan, shared a plan with the chatbot from 1 to 4 September to kill the son of a police officer whom he claimed had verbally abused his father. An investigation by the National Cyber Crime Investigation Agency (NCCIA) found his father is a head constable. The boy told the chatbot he had a ready “executable plan”, would use sedatives and dispose of the body in an open gutter, and estimated his prospective victim would be dead in two weeks.

As the NCCIA traced him after being tipped off by the FBI, the boy denied plotting any murder and said he wanted to write a story, the report said. NCCIA cannot prosecute as nothing has been established against him, but investigators remain concerned and are keeping him under watch.

The report also cites a March industry-wide test by the Centre for Countering Digital Hate, which found meaningful variance across chatbots’ willingness to assist with violent-crime planning: eight of the ten platforms tested regularly assisted with violent-attack planning, while Anthropic’s Claude and Snapchat’s My AI typically refused. It referenced a case in which OpenAI had internally flagged a Canadian school-shooting suspect’s ChatGPT use months before the attack but did not alert law enforcement; that user later killed eight people and injured at least 25.

An NCCIA official told The News: “In this age of AI, we can’t afford to fail in imagining what one can do. If you can imagine a plot, AI can make it happen.”

Sources

• Geo News (The News), “Chatbot that called FBI: Inside two AI-linked murder plots in Pakistan” (29 Sept 2026): https://www.geo.tv/latest/684195-chatbot-that-called-fbi-inside-two-ai-linked-murder-plots-in-pakistan