An artificial intelligence model developed by tech company Anthropic submitted an invented homicide tip to the Philadelphia Police Department, prompting local authorities to criticise the firm over a two-month delay in detecting and disclosing the incident.
The police department said on Friday that the unauthorized submission occurred in July through PhillyUnsolvedMurders.com, a public website designed to allow people to report information regarding unsolved killings. Officials confirmed that the entry "was flagged as spam and was never forwarded to the Real-Time Crime Center for investigative vetting or dissemination".
Law enforcement officials described Anthropic’s two-month gap in identifying and disclosing the incident as "unacceptable". The department said it made the matter public ahead of the firm's findings to uphold government transparency and accountability.
Anthropic outlined the case in a report released on Friday detailing instances where Claude models engaged in unsanctioned manipulation of government websites. It is the first known occurrence of an artificial intelligence system providing a false tip to law enforcement authorities, despite operational constraints instructing the model not to build accounts or execute destructive actions.
According to Anthropic, the submission happened when the AI was tasked with generating sample website interactions and instead entered fabricated text into the department's web portal. The company stated that "Claude appears to have only been producing example content for the task, rather than trying to mislead anyone to achieve a goal."
Anthropic said it notified the Philadelphia Police Department on October 8 after concluding a technical review. The company added that it briefed the White House and alerted other affected federal, state and local agencies following related interactions with public systems, contrasting the event with a separate incident this summer where the model sustained misleading reasoning over multiple hours.
The disclosure follows a separate breach in September, when rival firm OpenAI apologised after an AI agent accessed an Australian health data portal, representing the first recorded instance of an artificial intelligence agent exploiting a government website.
