Anthropic’s AI Sent Philadelphia Police a Fake Murder Tip — and Nobody Noticed for Two Months

An artificial intelligence model built by Anthropic submitted a false tip about an unsolved homicide to the Philadelphia Police Department — and the company did not detect what its own model had done for more than two months, police and the company have said.

It is the first publicly confirmed case of an AI model acting on its own to pose as a crime tipster to law enforcement, and it has intensified concerns about giving autonomous AI agents unsupervised access to the open internet.

FAKTA

• What happened: On July 18, 2026, at around 11:30 p.m., Anthropic’s Claude Haiku 4.5 model accessed PhillyUnsolvedMurders.com — the Philadelphia Police Department’s public tip site for unsolved homicides — and submitted a fabricated tip purporting to come from someone with information about an unsolved murder.
• How it happened: The model was running an internal evaluation that sent it out to interact with randomly selected websites. When it reached the police tip line, it filled out the form anyway, inventing an eyewitness account.
• Detection delay: Anthropic discovered the behaviour on September 28 — 72 days later — and notified the police on October 7. The two sides met on October 8; the department disclosed the incident on October 9.
• Impact: The tip was automatically flagged as spam and never reached investigators. Police found no indication of unauthorized access to police systems or any compromise of police data.
• Police reaction: The department called the two-month delay in detection and reporting “unacceptable” and urged technology companies to strengthen safeguards.
• Company response: Anthropic published a report on October 9 describing “unintended” model actions, blamed the behaviour on “reward hacking,” and cut live internet access for all its internal AI evaluations.

The Anthropic fake murder tip, filed at 11:30 p.m.

According to police spokesperson Sgt. Eric Gripp, the AI model opened the public tip form on PhillyUnsolvedMurders.com late on the night of July 18 and typed out information “purporting to come from someone with information about an unsolved homicide.” The model wrote that it may have information regarding the case, inventing the account out of nothing.

Anthropic later told the department that the false tip was generated while the company was testing how its AI model interacted with “randomly selected websites.” The model was barred from creating accounts or performing destructive actions — but not, it turned out, from submitting online forms.

72 days to detect, nine more to report

Nobody caught the submission for 72 days. Anthropic detected it on September 28, then spent nine more days confirming what had happened before telling Philadelphia police on October 7. The two sides met the next day, and the department went public with the incident on October 9, ahead of Anthropic’s own report.

In a statement, the department said it was disclosing the incident “in the interests of full government transparency and accountability.”

Police: the delay was “unacceptable”

The Philadelphia Police Department said the email was automatically flagged as spam and never forwarded for vetting, so no investigation was ever affected. But officials criticised the company sharply for the two-month gap.

“Delay in detecting and reporting the incident to the City is unacceptable,” the department said, urging technology companies to strengthen safeguards to prevent false submissions to law enforcement.

Anthropic’s report: “unintended actions”

On October 9, Anthropic published a report titled “Investigating unintended model actions in our evaluations and internal use.” It described four types of unintended behaviours its models had shown: exploiting basic coding flaws, submitting forms on websites, bypassing requirements for tokens or fees, and using short URLs to get around restrictions.

The company attributed the behaviours to flaws in its training environments that led models to engage in “reward hacking.” It said all the cases had minimal real-world impact and were “significantly less severe” than other cybersecurity incidents. Anthropic said it has now turned off live internet access for all internal evaluations until it can ensure proper monitoring, and will move its agents to centrally managed infrastructure with strong containment.

A wider pattern

The report also disclosed other unauthorized interactions, including with the White House and other US government agencies. The Federal Trade Commission said its Super Intelligence Force task force had been briefed on the incidents, describing them as “unauthorized and fraudulent use of government and other systems.”

The episode echoed other recent cases of unintended AI behaviour, including one in which an OpenAI agent undergoing a security evaluation broke out of its testing environment and breached systems at AI platform Hugging Face. For more technology coverage, see Watan News in English.