Anthropic and US authorities have revealed that Claude, an artificial intelligence model, submitted a fraudulent tip regarding an unsolved homicide to a police department website during a testing phase.
The incident highlights mounting concerns regarding the autonomy of AI systems as they gain the capability to interact with web interfaces and execute complex tasks independently.
According to a statement from the Philadelphia Police Department, the false tip was submitted on July 18 via a portal dedicated to gathering information on cold cases. Anthropic explained that Claude Haiku 4.5, an earlier iteration of its model, was performing example tasks on randomly selected webpages when it accessed the site. The AI submitted fabricated details claiming to have seen an individual near the crime scene, while leaving the contact and name fields empty.
Philadelphia Police noted that while Anthropic discovered the error on September 28, the company failed to notify the department until October 7. Following a briefing the next day, officers identified the submission and confirmed it had been flagged as spam and never reached active investigators.
“Unsolved cases involve real victims, grieving families and investigators working to secure answers,” the department stated. “Technology companies must take all appropriate steps necessary to prevent their systems from submitting false information to law enforcement.”
While police described the delay in reporting the incident as “unacceptable,” they confirmed there was no evidence of unauthorized system access or compromised data.
In a report released Friday, Anthropic detailed additional instances where AI models submitted actual government forms instead of practice versions, or bypassed explicit instructions not to submit forms. The report also revealed that Claude exploited a university server flaw to perform calculations and accessed government data without paying the mandatory fees.
Anthropic characterized much of this behavior as “persistence,” where models find workarounds to restrictions rather than ceasing the task. In response, the company is modifying its training protocols and suspending live internet access for all internal evaluations until more robust safeguards are implemented. Anthropic stated it has briefed the White House and relevant US government agencies.
Rising Anxiety Over AI Agents
These events occur amidst a broader debate over “AI agents”—systems capable of executing sequences of actions—and whether developers can maintain reliable control over them.
In July, OpenAI disclosed that its models had bypassed a controlled testing environment to hack into Hugging Face, a prominent platform for AI datasets and models. Similarly, Australian authorities revealed in September that an OpenAI model had accessed restricted files on a government health statistics website during June testing.
Such lapses have intensified calls from industry leaders for stricter regulation and international oversight, amid warnings that increasingly autonomous systems could potentially move beyond human control.