Florida Woman Arrested After Anthropic Flags AI Threat

A Southwest Florida woman faces a felony charge after artificial intelligence startup Anthropic detected threats in her chat history and reported her directly to local law enforcement.

Carli Michelle Heller, of Bonita Springs, was taken into custody by Lee County deputies following an investigation triggered not by a human tipster or social media post, but by internal safety monitors overseeing Anthropic’s Claude chatbot. According to arrest records, Heller entered prompts on Sept. 26 stating she planned to “shoot up” the Lee County Sheriff’s Office. A subsequent message the following day indicated she had acquired a firearm.

Automated filters embedded within Claude flagged the phrasing, routing the transcript to Anthropic’s human trust and safety personnel. The review team assessed the threat as severe and contacted authorities in Florida. Deputies dispatched to Heller’s residence arrested her without incident on a charge of making a written threat of violence. During questioning, Heller told investigators she treated the generative AI interface like a private diary.

The case underscores a stark legal reality that privacy advocates have repeatedly stressed: consumer AI platforms do not enjoy the legal confidentiality of a personal journal, a therapist, or an attorney. Under the legal third-party doctrine, information voluntarily transmitted to a commercial server forfeits Fourth Amendment protections against warrantless corporate disclosure.

Anthropic explicitly reserves the authority to inspect user communications. Its consumer terms grant the company sole discretion to turn over prompts, outputs, or identifying data to police. The company’s privacy documentation specifies that disclosures occur whenever safety teams possess a good-faith belief that sharing data is reasonably necessary to avert death or serious physical harm. Users who opt out of having their chats train Anthropic’s foundation models still remain subject to system retention policies, which store transcripts for 90 days for trust and security audits.

Flagged records serve a secondary operational function as well: Anthropic feeds confirmed violations back into internal safety pipelines to tune its machine-learning classifiers against violent rhetoric and harmful content.

The Florida intervention represents part of an increasingly aggressive posture by frontier AI labs wrestling with their liability surrounding violent crimes. In September, Anthropic alerted San Francisco authorities after a subscriber claimed to have acquired an assault rifle to target CEO Dario Amodei. While that user avoided immediate charges after claiming the post was satire, the company defended the escalation as a proper test of its defensive protocols.

Other industry leaders operate under similar liability burdens. Following the deadly February shooting in Tumbler Ridge, British Columbia, OpenAI disclosed that it had debated contacting Canadian authorities months earlier regarding the shooter’s suspicious prompts before ultimately deciding the interactions fell short of demonstrating an actionable, imminent threat. OpenAI subsequently overhauled its safety routing to ensure any direct threats of physical violence funnel quickly to human moderators empowered to contact police.

Heller remains scheduled for an initial court appearance in Lee County in November.