Dear Diary, Rat Me Out to the Cops

Anthropic called the cops on one of its users. With good reason: investigators say Carli Michelle Heller, a 30-year-old Florida woman, told Claude she was going to “shoot up” the Lee County Sheriff’s Office. She wrote that on September 26, then mentioned a new gun the next day. Heller later told investigators she used the chatbot as a personal diary. Apparently, this one forwards entries to the cops.

The messages went through an automated safety check, then a human review, before Anthropic notified law enforcement. Heller was arrested and charged with making a written or electronic threat of a mass shooting or terrorist act, a felony under Florida law. The chat is now part of the case. She didn’t even have to send the threat to the sheriff herself.

Anthropic’s policy allows emergency disclosures without waiting for the usual legal process. If the company believes someone faces imminent physical harm or death, and sharing information immediately could prevent it, it can hand over the data. Here, the service raised the alarm itself. A conversation that seemed to have two participants suddenly had several more.

“We will make an exception to this policy if we believe there is an emergency that may result in imminent physical harm or death” — Anthropic’s government-request policy

A chat window can feel like a room with the door shut. The bot has no friends to gossip with. It listens patiently, never looks bored, never seems annoyed. Easy enough to start treating it as a place to say things nobody else will hear. Except the company owns the room. Its screening systems and employees can read the messages and decide who else gets to see them. Someone sits down for a private vent and ends up supplying material for a police investigation.

That friendly interface gives very little away about where user content can end up. We’ve already covered OpenAI’s agents, which posted images uploaded by ChatGPT users to third-party sites. OpenAI said the transfers should never have happened and worked to get the files removed. Heller’s messages reached the police by design: Anthropic’s safety team made a deliberate decision to report them.

Reporting a threat to shoot up an office is a perfectly reasonable call. The same incident also exposes the cost of casually treating a chatbot as a diary, a therapist and a confessional. The interface supplies the intimacy. The company decides who gets access to the confessions. That endlessly patient listener comes with a safety department and a line to law enforcement. Heller thought she was keeping a diary. She was writing her own tip-off.