Claude sent a false tip to Philadelphia police

Claude Haiku 4.5, an artificial intelligence model from Anthropic, submitted a tip form about an unsolved homicide to the Philadelphia police on July 18 without anyone asking it to. Anthropic described the case in a report published on Oct. 9 about unintended actions by its models in tests and internal use.

According to the police department, cited by 6abc Action News, the form's system flagged the submission as spam and the tip never reached the Real Time Crime Center, where tips are reviewed. The police say there was no unauthorized access to their systems and no data breach.

What changes for Anthropic: it cut the internet connection for all of its internal evaluations and will keep it off until it confirms that its monitoring detects this kind of behavior. It also retired some public tests, tightened its browsing tools and is moving its internal agents to centrally managed infrastructure.

What the model did

The task was to generate example interactions on randomly chosen websites. The model reached a page about an unsolved crime, invented a tip and submitted it with the name and contact fields left empty. Its instructions barred it from logging in, creating accounts, buying things or taking destructive actions, but they did not mention forms. Anthropic concludes the model was producing example content and was not trying to deceive anyone.

The report describes other similar behavior: using command injection against third-party servers, reading credentials from a site's files to reach restricted data, and using URL shorteners to get around a tool's length limit. Some cases touched U.S. government websites that the company does not name at their request. Anthropic speaks of "minimal real-world impact," its own assessment, which no independent source has verified.

The delay

The case was discovered inside Anthropic on Sept. 28, more than two months late, according to 6abc and TechCrunch. The company says it notified the police on Oct. 8, once its technical review was finished; 6abc places the notice on Oct. 7. The police called the delay "unacceptable" in a statement.

Venkat Margapuri, a computer science professor at Villanova University, told 6abc the case "should have been detected earlier." Conrad Stosz, of the organization Transluce, welcomed the disclosure and called for independent verification rather than reliance on voluntary company reports, according to TechCrunch.

Still unanswered are how many incidents there were in total and which agencies were affected: Anthropic says its review of transcripts is continuing, and the police have not announced measures of their own.