What happened

AI company Anthropic reported 30-year-old Carli Michelle Heller of Bonita Springs, Florida, to law enforcement over threatening messages written in its Claude chatbot. The company reached that decision after the chatbot conversations went through its safety systems. The Verge reported the story on October 5, 2026.

The chain of events began on September 26, 2026: that day, Heller wrote in the Claude chat about her intent to attack the Lee County Sheriff's Office. A day later, on September 27, at about 1:07 AM, she left an additional message saying she had a new gun, calling it her "last chance." Four days passed between the first message and the arrest on September 30. Both messages were written in the Claude chat.

The company's automated safety systems detected these messages. Once the suspicious content was identified, the conversation was automatically escalated to a human review team. The team examined the messages and assessed them as a credible threat — only after that assessment did Anthropic formally contact law enforcement.

Deputies from the Lee County Sheriff's Office came to Heller's home on September 30, 2026, and detained her without incident.

Sheriff Carmine Marceno said Heller told officials she had been using Claude as a diary. Marceno reminded users they are "never truly anonymous" in AI chatrooms and searches, stressing that threats are taken seriously.

This is reportedly at least the third case since August in which AI chat correspondence has been handed to police. Anthropic did not immediately respond to a request for comment.

How the safety systems worked

In this case, three stages of Anthropic's safety chain engaged. The first stage is automated detection: the system screens user messages and singles out suspicious content. The second stage is human review: the conversation is handed to a team of people, which assesses the credibility of the threat. The third stage is formal referral: if the threat is judged credible, the company notifies law enforcement.

This sequence is set in the company's internal policy: automated detection is not a final verdict; the final decision is made by the human team.

Chats flagged for review are also used to improve the detection of harmful content — that is stated in the company's policy.

Heller herself told officials she had been using Claude as a diary:

Heller told officials she had been using Claude as a "diary." — According to The Verge

When the company discloses data

Under Anthropic's policy, the company's safety team may disclose user data only in limited emergencies: when sharing the information is deemed necessary to prevent death or serious physical injury.

This criterion is a clearly defined high bar in the policy. The right to disclose belongs specifically to the safety team: in emergencies, that team decides whether to disclose user data. Referral to law enforcement happens only when the threat has been assessed as credible by the human team.

Anthropic's transparency policy states that law enforcement data requests are reviewed "in accordance with applicable laws while protecting user privacy."

Heller was detained on September 30, 2026. She was charged with making written threats of violence — under Florida law a second-degree felony. The charge is based on Heller's correspondence in the Claude chat. The detention proceeded without incident; a court hearing is scheduled for November. After the detention, the investigation was taken over by an intelligence detective with the Lee County Sheriff's Office.

More details are in The Verge's report.