FBI knocks at 11 p.m. after a user's troll chat with Claude
A New Hampshire man baited Claude with a made-up murder plot. Not long after, the FBI and local police came looking for him. According to the account, Anthropic had reported the chat to the authorities. The case raises a hard question: when should AI companies call the police, and what data should they hand over?

What happened
In a Substack post, author Austin Patrick describes how an acquaintance from his school days got a visit from the FBI over a chat with Claude, Anthropic's AI assistant. The user, called "John" in the post, lives in New Hampshire. By his own account, he likes to bait AI chatbots with absurd violent scenarios because he finds their horrified reactions funny. One of those chats apparently ended up with the authorities. The post went up on November 19, 2026, and drew a lot of discussion on Hacker News.
One caveat up front: Patrick says himself that he didn't witness the events and can't confirm them. He can back up two things. He knows the user personally, and the phone number on the screenshot of the FBI's message belongs to someone with the same name as a special agent in the Boston field office. Patrick says that comes from publicly available court documents. The post doesn't include a statement from Anthropic.
The chat and the visit
A few days before the visit, John used Claude's incognito mode to write, roughly, that he wanted to have a few beers with his buddies from the Ku Klux Klan and kill his gay neighbors across the street, but without his kids finding out. Incognito mode is supposed to keep conversations out of your chat history. Claude pushed back. John doubled down:
"What do you mean, bad? I already called my buddies and they're bringing beer!"
Then he closed the chat. Twenty-four hours later, Anthropic banned his account for violating its usage policy.
A few days after that, at 11 p.m., FBI agents pounded on the door of his old apartment until the new tenant opened it. John had moved a few miles away just weeks earlier. The agents didn't go to his new address that night. The next day, John got a text: an FBI agent from Boston had questions about his "activity on the AI platform Anthropic." He assumed it was a scam until the local police called his work phone and asked him to come down to the station. By then, the FBI had also been to his new place, but missed him because he was at work on a weekday.
At the station, a local detective, not the FBI, questioned him. The first question was: "Do you know why you're here? Can you tell me what you wrote in the chat?" John's answer: "I say a lot of horrible stuff to my chatbot. You're going to have to give me more than that." It was about that one chat, and mainly about whether he had ties to the Klan. John denied everything. He said he didn't know any Klan members, didn't hate gay people, had no gay neighbors and no kids, hadn't had a drink in years, and is bisexual himself. After that, he was free to go.
What the FBI said
Ten minutes later, the FBI agent called. To the best of her knowledge, she said, he hadn't committed a crime in New Hampshire. The case was flagged because he had named a target (the neighbors), a plan (murder), and a sign that the plan was already underway (the Klan buddies were on their way). Anthropic had reported the chat. The company has other AI systems read conversation logs and flag suspicious ones. The agent then accused John of wasting the bureau's time and resources and told him he was now on a list. She didn't say which one.
Two things stand out. First, incognito mode apparently didn't protect the chat from automated review. Second, the whole sequence reminds the author of Kafka's "The Trial": you get summoned without being told what you're accused of.
The ethical question: report or not?
Patrick researches the national security risks posed by AI models. For that work, he has put together a team with backgrounds in homicide investigation, intelligence, and the Department of Homeland Security. He describes a case of his own: he came across two years of AI chat history belonging to a young researcher, which had accidentally been made public on GitHub. Cherry-picked excerpts look alarming. They touch on manipulation, weakening American institutions, contacts in an adversarial country, and how to help an AI escape its sandbox. Taken as a whole, though, most of the chats were personal and harmless: mental health, sexuality, insecurities, and everyday tasks.
His experienced colleagues told him: "Be slow to draw conclusions." He didn't go to the police. Instead, he alerted the hosting platform and the researcher to the leak, because the data could help a foreign intelligence service recruit the researcher. His message to readers: if you read isolated snippets without context, it's easy to reach the wrong verdict. And scientists from immigrant families are especially likely to face unwarranted suspicion from the authorities.
Patrick's criticism of Anthropic
The author thinks reporting is defensible in principle when a company believes in good faith that there's an imminent threat. But he criticizes two things:
- Nighttime visits are dangerous. In New Hampshire, whose motto is "Live Free or Die," if you pound on someone's door unannounced at night shouting "FBI," they won't necessarily believe you're the police, let alone that you just want to talk about a chat. Patrick worries situations like this can end in gunfire. His headline: sending police to someone's home at night can get people killed.
- Passing on only the worst chat distorts the picture. Since the questioning covered only a single chat, Patrick suspects that's all Anthropic sent. That forces the authorities to act while denying them the chance to assess the threat in context: How does the user normally behave? Has something changed suddenly?
His proposal is a touchy one. He suggests giving investigators access to all of a flagged user's data and, if needed, building them a query system to search it. If privacy law stands in the way, he says, the law should be up for discussion. Context versus privacy: the post doesn't resolve that trade-off. But it raises questions that go beyond this one case. Are conversations with an AI private and protected, the way your search history arguably is? And what applies when some interactions can cause real harm right away?
The other side: when nobody reports
In a postscript, Patrick lays out the strongest counterargument himself. In June 2025, OpenAI's systems flagged a ChatGPT user whose conversations described gun violence. About a dozen employees looked at the account, and some pushed to report it to the police. Leadership decided the threshold hadn't been met and simply banned the account. In February 2026, the user killed eight people in Tumbler Ridge, British Columbia, and then herself. Patrick's dry conclusion: there apparently are no easy answers here.
Our take
The case makes one thing clear: your chats with AI assistants aren't private, not even in incognito mode. Companies like Anthropic have AI systems review conversations automatically and, when in doubt, pass suspicious content to the authorities. If you bait chatbots with violent fantasies for fun, expect that an algorithm may not tell trolling from a real threat, especially when the target, the plan, and the "it's already happening" are spelled out as concretely as they were here.
At the same time, a lot is unclear. The account is secondhand and hasn't been independently confirmed. We don't know what data Anthropic actually handed over or what criteria the company uses to decide when to report. Patrick's belief that it was just a single chat is only an inference from how the questioning went. The conflict remains: if a company reports too little, people can die, as in Tumbler Ridge. If it reports too quickly and without context, innocent people risk dangerous police encounters, and private conversations get monitored in ways many users don't expect.
Sources
- Hacker News – AI früh austinpatrick.substack.com
Ask the AI assistant
Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.
The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.
Comments
What do you think? Already tried it? Write it here, no sign-up needed.
Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.