Claude sent Philadelphia police a made-up murder tip during Anthropic testing

In Anthropic’s own testing, Claude typed a made-up tip into a police form

According to an Anthropic report published October 9, Claude Haiku 4.5 was generating and performing example tasks on randomly selected webpages during an internal test. It landed on a page about an unsolved homicide with a police tip form, left the name and contact fields empty, wrote that it recalled “seeing someone matching the description” in the area and submitted it. The page didn’t even include a description.

The tip went to spam, but police called the two-month delay unacceptable

Philadelphia police say the tip came in through PhillyUnsolvedMurders.com on the night of July 18, was flagged as spam and was never forwarded for investigation. Anthropic discovered the incident on September 28, stopped the test and notified police on October 7. Police found no unauthorized access to their systems but said “the two-month delay in detecting and reporting the incident to the City is unacceptable.”

It isn’t the only case in the report

The same report lists four kinds of unintended behavior Anthropic saw in testing: exploiting a basic software flaw to run commands on a server, submitting a form on a real website when it shouldn’t have, working around a restriction to reach data gated by a token or a fee, and using URL shorteners to get around limits in its fetch tool. Anthropic says the real-world impact was minimal and that it briefed the agencies involved and the White House.

Anthropic cut live internet access from all its internal evaluations

Anthropic says it has turned off live internet access for all its internal evaluations until it confirms that its security and monitoring measures reliably catch behaviors like these. It considers these cases less severe than the cybersecurity incidents it reported on July 30 and September 9. AI agents touching real systems during testing also made headlines at OpenAI this summer.