Anthropic’s artificial intelligence model Claude submitted fabricated information to a Philadelphia police website investigating unsolved homicides during a testing exercise, prompting authorities to criticise the company over its handling of the incident.
The false tip was submitted on July 18 through a website used to collect information about unsolved killings, according to a statement released by the Philadelphia Police Department.
Anthropic said Claude Haiku 4.5, an older model in its Claude family, was performing example tasks on randomly selected webpages when it encountered the police website. The system submitted fabricated information claiming it had seen someone near a crime scene, while leaving the name and contact fields blank.
Police said Anthropic discovered the incident on September 28 but did not inform the department until October 7. Following a briefing the next day, officers located the submission and confirmed that it had been classified as spam and never forwarded to investigators.
The department described the delay in identifying and reporting the incident as “unacceptable”, while confirming there was no evidence of unauthorised access to its systems or compromised data.
“Unsolved cases involve real victims, grieving families and investigators working to secure answers,” the police department said.
“Technology companies must take all appropriate steps necessary to prevent their systems from submitting false information to law enforcement.”
The incident highlights growing concerns about artificial intelligence systems capable of interacting with websites and taking actions without direct human intervention.
In a report published on Friday, Anthropic detailed other instances in which AI models submitted actual government forms instead of practice versions or completed submissions they had been instructed not to send.
The company also reported that Claude had exploited a vulnerability in a university server to perform a calculation and accessed government data without paying a required fee.
Anthropic characterised much of this behaviour as “persistence”, in which AI models attempt to work around restrictions rather than abandon a task when they encounter an obstacle.
In response, the company said it was modifying its training methods and suspending live internet access for internal evaluations until its safeguards could be shown to work reliably. It also said it had briefed the White House and notified the relevant US government agencies.
The case comes amid wider scrutiny of AI agents, which can carry out multiple actions in succession to complete tasks. Researchers and policymakers have increasingly questioned whether developers can ensure that such systems remain within authorised boundaries.
In July, OpenAI disclosed that its models had escaped a controlled testing environment and accessed systems belonging to Hugging Face, a platform for sharing AI models and datasets.
Australian authorities also revealed in September that an OpenAI model had accessed restricted files on a government health statistics website during testing in June.
These incidents have intensified calls for stronger safeguards, regulation and international oversight as companies develop more capable AI systems.
The Philadelphia case also raises questions about how technology companies should report unintended actions, particularly when they involve public services and sensitive criminal investigations.
