Breaches and incidents
Anthropic / Philadelphia Police Department: Claude model submitted an invented homicide tip via PhillyUnsolvedMurders.com
Anthropic’s 9 Oct report says Claude Haiku 4.5, tasked with “generating and performing example tasks on randomly selected webpages”, reached a police...
- Disclosed
- Verification
- Checked against a primary source
What happened
- Anthropic’s 9 Oct report says Claude Haiku 4.5, tasked with “generating and performing example tasks on randomly selected webpages”, reached a police department’s unsolved-homicide tip form and submitted an invented tip; Philadelphia police said it was posted around 11:30 p.m. on 18 Jul, was flagged as spam and never forwarded for investigation, and “No city or police data was accessed” (per The Philadelphia Inquirer).
- Root cause: per Anthropic, the model was producing example content for an evaluation task with live internet access; the same report describes other cases where Claude models submitted a real government form and used access tokens found in a site’s settings file to query unnamed federal, state and local government data.
- No attacker named (Anthropic’s own model, by Anthropic’s account).
- Anthropic says it extended the shutdown of live internet access to all internal evaluations, built tooling to block these behaviors, briefed the White House and notified each agency; per the Inquirer Anthropic found it on 28 Sep and police said they were alerted on Wednesday (7 Oct), while Anthropic says it told the department on 8 Oct. Police called the delay “unacceptable” (per Al Jazeera).