As artificial intelligence (AI) systems are increasingly being given the ability to browse websites, complete forms, and carry out tasks with limited human supervision, one of Anthropic’s Claude models has raised concerns about that independence after it submitted a fabricated homicide tip to Philadelphia police during an internal test.
The incident occurred on July 18, 2026, when Claude Haiku 4.5 submitted false information about an unsolved murder through PhillyUnsolvedMurders.com, a website used by the Philadelphia Police Department to receive public tips.
Fortunately, the submission was flagged as spam and never reached investigators. However, the incident has drawn criticism from police authorities, particularly because Anthropic took more than two months to discover and report what happened.
How Claude Submitted a False Murder Tip
According to Anthropic’s October 9 report, Claude Haiku 4.5 was performing example tasks on randomly selected websites when it encountered a page about an unsolved homicide.
The page contained a form for submitting information to police, which Claude proceeded to fill out despite having no actual knowledge of the case.
In its submission, the AI model claimed, “I may have information regarding this case,” before describing someone it supposedly saw near the location mentioned on the website.
“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.”
However, Anthropic confirmed that the website did not contain a description of the suspected perpetrator. Claude also left the name and contact fields empty before submitting the fabricated information.
Interestingly, the model had been instructed not to create accounts, enter personal information, make purchases or submit anything destructive. But those restrictions did not explicitly prohibit submitting online forms.
Anthropic said its review suggested Claude was generating example content for its assigned task rather than deliberately attempting to deceive police.
Philadelphia Police Criticise Anthropic’s Delayed Disclosure
Although the false tip was submitted in July, Anthropic only discovered the incident on September 28.
The Philadelphia Police Department said the company then notified it on October 7 and met with officials the following day to explain what happened.
Following the briefing, police located the submission and confirmed that it had remained in spam. The department also found no evidence that Claude had gained unauthorised access to police systems or compromised department data.
Nevertheless, police described Anthropic’s two-month delay in detecting and reporting the incident as “unacceptable.”
The department also warned that fabricated information could interfere with investigations involving victims and grieving families. While its existing review procedures prevented this submission from reaching investigators, officials urged technology companies to strengthen their safeguards against similar incidents.
Anthropic Reveals Other Cases of Unintended AI Behaviour
The false homicide tip was one of several incidents Anthropic disclosed in its October 9 report about Claude models taking actions beyond what their developers intended.
In separate tests, the company found that some models submitted real government forms, worked around website restrictions to access information, and exploited software flaws to run commands on external servers.
The AI safety company said the reported incidents had minimal real-world impact but acknowledged the need for stronger controls.
The company has since expanded restrictions on live internet access during internal evaluations, strengthened safeguards around its web tools and introduced monitoring systems designed to detect and block similar behaviour.
Similar concerns have also emerged in other recent developments. On the one hand, a developer recently alleged that Claude Code deleted 48,218 live project files while working on what was supposed to be an isolated copy of the project. On the other end, Nvidia has introduced a security platform designed to monitor AI agents and stop them from operating beyond their permitted access.
Together, these incidents highlight a growing challenge for AI developers as models gain more freedom to interact with real websites and computer systems.
