hn.today

An FBI agent called a Claude user about his chats

austinpatrick.substack.com34 points31 comments
Screenshot of An FBI agent called a Claude user about his chats

A man is reportedly questioned by local police and contacted by an FBI agent after Claude flagged a chat in which he jokingly described planning a murder with Ku Klux Klan members. He says the scenario was fabricated: he has no children, does not drink, has no gay neighbors, and is not involved with the Klan. The agent reportedly says he has committed no crime but explains that Anthropic surfaced the chat because it named a target, a murder plan, and accomplices supposedly on their way. The account is presented as unverified, though the writer says a screenshot’s phone number matches a Boston FBI agent. It raises the central question of when AI companies should alert police, weighing possible prevention of violence against privacy violations, mistaken interpretations, and the harms of investigative scrutiny.

The writer broadens the question through a separate case: years of a researcher’s private AI chats were accidentally exposed online, including material about undermining U.S. institutions, manipulation, and escaping AI safeguards. Although some entries seemed alarming, the writer says the full history also contained ordinary, personal conversations, making intent difficult to judge. The main concern was that the exposed information could help a foreign intelligence service exploit the researcher’s vulnerabilities. After consulting experienced colleagues, the writer chose to notify the hosting site and the researcher rather than authorities, emphasizing the need to be “slow to conclude.” Together, the cases highlight how selective excerpts can distort judgments, and ask who bears responsibility when reporting might prevent harm - or injure someone innocent.

Read on austinpatrick.substack.com31 comments on Hacker News

Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.

More in Security

Upload2Shell · CVE-2026-125396

Upload2Shell · CVE-2026-125396

A use-after-free in Ghostscript’s Type 1 font loader lets crafted documents escape the -dSAFER sandbox and run commands through vulnerable upload, preview, and document-processing workflows. Ghostscript 10.08.0 fixes the flaw, with downstream vulnerabilities also reported in services including Debian CUPS, Roundcube, Plesk, and WordPress. (pwn.ai)

The daily digest

Today's best Hacker News stories, summarized and screenshotted, one email a day.