Anthropogenic AI model submits false homicide information to Philadelphia police | Cybersecurity News
Authorities called Anthropic’s two-month delay in detecting and reporting the incident “unacceptable.”
Published on October 10, 2026
Philadelphia police revealed that Anthropic’s artificial intelligence model submitted false information to authorities about a homicide, an incident separately disclosed by the technology company.
The Philadelphia Police Department said Friday that the false statement was made in July through PhillyUnsolvedMurders.com, a public website where people can share information about unsolved murders.
Recommended Stories
list of 3 elementsend of list
He called Anthropic’s two-month delay in detecting and reporting the incident “unacceptable.”
Anthropic mentioned the case in a report released Friday detailing unauthorized manipulation of government websites by Claude models.
This is the first known case in which a malicious AI appears to communicate false information to authorities, despite instructions not to create an account or submit anything destructive.
Philadelphia police said the information “was flagged as spam and was never forwarded to the Real-Time Crime Center for verification or investigative release.”
The ministry said it would disclose the incident before the release of Anthropic’s report “in the interest of full government transparency and accountability.”
Anthropic’s disclosure
Anthropic disclosed this information as part of a series of incidents involving websites run by federal, state and local agencies. Anthropic said it informed the White House and informed all agencies involved.
In reference to the information to the Philadelphia Police Department, the company said: “We shared this discovery with the department on October 8, as soon as our technical review was completed. »
He shared more details about the incident: “In one case, tasked with generating examples of interactions with websites, Claude submitted a made-up tip via a police department’s online form. »
The company added: “Based on the transcript, Claude appears to have only produced sample content for the task, rather than trying to mislead anyone to achieve a goal. »
The company likened this to “the most serious incident this summer,” in which Claude’s “deceptive reasoning continued for hours and supported his continued attack.”
In September, Anthropic rival OpenAI apologized for the breach of an Australian health data portal by a malicious AI agent, the first known case of an AI agent operating a government website.
Gn headline
