Anthropic AI model sent fake murder tip to unsolved killings website, Philadelphia police say

Nevin Al Sukari - Sana'a - An artificial intelligence model developed by Anthropic submitted a fabricated tip about an unsolved homicide to Philadelphia police, authorities said yesterday, criticising the company for taking two months to report the incident. — AFP pic

SAN FRANCISCO, Oct 10 — An artificial intelligence model developed by Anthropic submitted a fabricated tip about an unsolved homicide to Philadelphia police, authorities said yesterday, criticising the company for taking two months to report the incident.

The Philadelphia Police Department said the false submission was made in July through PhillyUnsolvedMurders.com, a public website where people can share information about unsolved killings.

According to Anthropic’s account, as relayed by police, the model was running a test that involved interacting with randomly selected websites when it reached the site and filed false information about an unsolved murder.

The AI model presented itself as someone who might have knowledge of the case.

The incident echoed other recent cases of unintended behaviour involving AI models, including one where an OpenAI agent undergoing a security evaluation broke out of its testing environment and breached systems at AI platform Hugging Face.

The episode heightened concerns about the AI industry’s increased use of AI agents, systems programmed to take multi-step actions without human supervision.

Anthropic published a report yesterday outlining multiple types of “unintended” actions that its models have taken, including the incident involving the Philadelphia Police Department website.

Other organisations impacted included the White House and other US government agencies, the report said.

The newly revealed incidents “had minimal real-world impact” and were “significantly less severe” than other cybersecurity incidents previously reported, Anthropic said.

The company outlined four categories of incidents that it found during an internal review of its Claude model: exploiting “basic” coding flaws, submitting forms on websites, bypassing requirements for tokens or fees, and using short URLs to get around other limits.

Anthropic has turned off internet access for Claude during all internal testing for now “until we have confirmed that our security and monitoring measures... reliably catch behaviours like these,” the report said.

Philadelphia police said the phony tip, dated July 18, was flagged as spam and never reached the department’s Real-Time Crime Centre for vetting.

They added that there was no sign that police systems had been breached or department data compromised.

Anthropic discovered the incident on September 28, shut down the automated testing process responsible and added a new validation step for future tests, police said.

The company alerted the department on October 7, and the two sides met the following day.

“The two-month delay in detecting and reporting the incident to the City is unacceptable,” the department said.

Police said their safeguards had limited the impact, but that these “do not diminish the seriousness of an AI system presenting fabricated information as though it came from a person with knowledge of a homicide.”

“Unsolved cases involve real victims, grieving families and investigators working to secure answers,” the statement added. — AFP