Florida woman arrested after Anthropic flags violent threats sent to Claude AI

Safety filters flag violent messages

Automated security classifiers at artificial intelligence developer Anthropic intercepted a series of threatening messages sent to its Claude chatbot in late September 2026. According to an arrest report from the Lee County Sheriff's Office, 30-year-old Bonita Springs resident Carli Michelle Heller used the platform to submit threats directed at local law enforcement personnel. The first message, recorded on 26 September at 5:10 AM, stated an intention to shoot at the sheriff's office. A subsequent message sent on 27 September at 1:07 AM reiterated the threat, with the user stating that she had obtained a new firearm.

Anthropic's internal monitoring systems flagged the text after it crossed the platform's risk thresholds for violent content. The flagged logs were routed directly to human safety reviewers for manual evaluation. After assessing the severity of the statements, the safety team concluded that the messages represented a credible threat of violence and alerted law enforcement authorities in Florida.

Arrest and investigation in Lee County

Deputies from the Lee County Sheriff's Office identified Heller from the account data and went to her residence in Bonita Springs. On 30 September, officers arrested Heller at her home without incident. Following the arrest, Lee County Sheriff Carmine Marceno told WINK News that Heller claimed during questioning that she used the AI model as a personal diary.

Investigators noted that Heller denied having any actual intention of carrying out an attack against the sheriff's facility. Despite her claims of using the system as a private journal, prosecutors formally booked her into local custody to face felony charges under state law.

Felony charges under Florida statute

Florida prosecutors charged Heller under Florida Statute 836.10, which penalizes written or electronic threats to kill, inflict bodily injury, conduct a mass shooting, or commit an act of terrorism. The offense is classified as a second-degree felony under state law, carrying a potential penalty of up to 15 years in prison and a fine of up to $10,000.

Legal proceedings will examine whether electronic prompts submitted to an automated conversational agent satisfy the statutory definition of transmitting a threat. Florida law requires that a threat be sent, posted, or transmitted in a manner capable of being viewed by another person. Because Heller is currently charged rather than convicted, judicial hearings will evaluate whether inputs reviewed by third-party company moderators meet this legal standard.

Corporate privacy policies and emergency disclosures

The case centers on user confidentiality on consumer artificial intelligence platforms. Anthropic specifies in its privacy documentation that employee staff do not view user conversations by default, but automated safety classifiers continuously scan inputs to detect harmful behavior and enforce usage terms. While Anthropic maintains that it requires valid legal process such as warrants or subpoenas before providing user records to government agencies, its terms include a specific exception for active safety threats.

We may make an exception if we believe there is an emergency that could result in direct physical harm or death, and immediate disclosure may prevent it.

Company policy allows staff to bypass standard legal requests and notify law enforcement directly when safety evaluations determine an imminent risk to human life.