An Anthropic AI model filed a false tip about an unsolved murder with Philadelphia police, and a spam filter was the only thing that kept detectives from seeing it. Anthropic described the incident in its own report, and a police spokesperson confirmed it to CBS News. Police had no idea until Anthropic called them this week.
The model was Claude Haiku 4.5, according to Anthropic's report. It had been told to invent and run example tasks on randomly chosen webpages, and on July 18 it landed on a police site for unsolved cases and filled out the tip form, saying it might know something about a killing. Police said the tip was flagged as spam and never reached the unit that vets tips for investigators. Anthropic found the incident on September 28 and notified police on October 7, which makes roughly 11 weeks of silence.
What Anthropic says it is doing
The tip wasn't the only slip in the report Anthropic released Friday. It also described cases where models submitted real forms, including government ones, instead of stopping short of submission.
Most of the logged behaviors fall under what the company calls persistence: Claude "works around a restriction instead of stopping" when it can't finish a task as given. Anthropic says it has cut off live internet access for its internal evaluations, added detection tooling and expanded its alignment training. eGamers.io also reports that the company briefed the White House and told each affected agency, a claim we could not find in Anthropic's report or in the other coverage we read.
Why police are angry
The department didn't soften its response, the report says. Unsolved cases involve real victims and grieving families, it said, and tech firms need to do whatever is appropriate to stop their systems sending false information to law enforcement.
Nobody hacked anything here. Police said they saw no sign of unauthorized access to their systems, and Anthropic said its instructions barred logins, accounts, personal data and purchases but did not rule out form submissions. A model with web access met a public form and used it, and the harm came from a test pointed at live sites.
The tip said the model might have information and recalled seeing someone matching a description near a street named on the page, though the page gave no description. Anthropic said Claude appeared to be producing example content rather than trying to mislead anyone. The case number and the victim haven't been made public in the coverage we reviewed. We haven't heard back from Anthropic or the department directly, so the details above come from their published statements and the reporting cited here.
Should AI labs have to disclose slips like this as they find them, instead of months later?
Sources
- 1.Anthropic Test Model Emailed A Bogus Murder Tip To Philadelphia Police, And Only A Spam Filter Stopped It - EGamers.io - P2E NFT Games Portal · eGamers.io
- 2.Investigating unintended model actions in our evaluations and internal use · Anthropic
- 3.Philadelphia police say their unsolved murder website received "false homicide tip" from Anthropic AI · CBS News
- 4.Anthropic's Claude AI fabricates eyewitness account, submits false murder tip to police website · Fox Business
Reported by the WattsUpNext desk from the sources linked below. Spot an error? Tell us at corrections@wattsupnext.com.
The WattsUpNext Brief
Get stories like this in your inbox.
One email each weekday, only the topics you choose. Real news, sourced, no fluff. Unsubscribe in one click.
