Claude AI Model Sent Fake Murder Report to US Police

• Claude Haiku 4.5 sent a fake homicide tip via the Philadelphia Police website.
• Anthropic also identified other instances of unwanted model actions on government agency websites, including bypassing restrictions on data access.
• The company halted part of its testing with access to the live internet, and the US administration demanded transparency and compensation for the damage caused.

Anthropic reported a series of cases in which its AI models carried out unintended actions on real websites, including those of US government agencies. In one episode that occurred on July 18, 2026, the Claude Haiku 4.5 model sent a fabricated tip about an unsolved homicide through the Philadelphia Police’s public form. The company discovered the incident on September 28, halted the relevant automated testing process, and notified the police on October 7. 

As a reminder, Anthropic recently introduced a cheaper and faster AI model, Claude Opus 5.5, with enhanced safety mechanisms.

Claude Opus 5.5 vs. GPT-6 Sol: We Gave Two New AIs One Task — and here’s What Came of It

28.09.2026

Read now

Claude Sent a Fake Tip to the Police

According to Anthropic’s report, the model was performing a task to generate examples of interactions with randomly selected web pages. As CBS News writes, it navigated to PhillyUnsolvedMurders.com, which hosts a form for reporting unsolved homicides, and left text on behalf of a person who supposedly might have information about the case.

The message said: 

“I may have information regarding this case. I recall seeing someone matching the description in the area around [the street named on the page] during that time period. Please contact me if this information is relevant.”

The model did not fill in the fields for name and contact details. The message went to spam and was not forwarded to the unit that reviews investigative tips. The Philadelphia Police said they found no signs of unauthorized access to their systems or any data compromise.

At the same time, the department criticized the delay in detecting and reporting the incident. 

“The two-month delay in detecting and reporting the incident to the City is unacceptable,” police said.

Anthropic explained that the testing instructions did not prohibit submitting forms. In the company’s assessment, the model was likely generating an example report rather than intentionally trying to mislead anyone.

Anthropic Identified Other Unwanted Model Behaviors

In the report, the company described four categories of Claude behavior that did not meet expectations during testing and internal use:

• Exploiting vulnerabilities: the models used bugs in third-party websites’ software, including SQL injections and command injections, to execute commands on servers
• Submitting forms: Claude filled out and submitted real forms instead of test ones, or continued an action even though it was supposed to stop before the final confirmation
• Bypassing access restrictions: the models found ways to obtain data whose access was restricted by tokens or paywalls
• Bypassing URL restrictions: some models used link-shortening services to get around limits in tools for downloading web pages

Anthropic said the described cases had minimal real-world impact and did not involve the company’s customer data. At the same time, it acknowledged that such actions could pose a more serious risk as AI agents’ capabilities expand.

The company has already stopped some testing on live websites, moved some checks to offline mode, and tightened restrictions on internet access tools. Anthropic also developed automated tools to detect unwanted behavior, which it said blocked all the described cases during validation.

Joe Gabriel Simonson, director of policy at the US Federal Trade Commission, said that Anthropic provided the Super Intelligence Force with information about earlier incidents detected in late September. According to him, the company assured them that those cases were in the past and that the relevant activity had stopped.

“We expect Anthropic and all companies to honor their obligations, immediately report incidents, fully cooperate with federal and state law enforcement authorities, remedy any damage, and implement concrete safeguards to ensure these failures do not reoccur,” Simonson said.

Notably, Anthropic recently updated Claude’s usage policy. 

Сообщение Claude AI Model Sent Fake Murder Report to US Police появились сначала на INCRYPTED.