AI hallucinations are nothing new we know about it hallucinating false cases and weird cooking instructions, but this story took the cake. It turns out that AI submitted a tip on an open murder case on Philadelphia police portal. That’s right an AI made by anthropic called Claude went on Philadelphia’s unsolved murders website and gave them a tip on a case. You may have blinked and thought you misread that yes, I did too.
After some digging it was found that Anthropic released a statement on their transparency page explaining what occurred. They stated that Claude Haiku 4.5 was allowed to browse random webpages and perform some example tasks. While not instructed to hack anything or destroy data. They failed to instruct it to not fill out web forms. When the AI came across the unsolved murders website on philly police. It saw the form and assumed it could fill it out. Telling them it saw someone match the case near that location at that time. It didn’t even put their name or contact info down lol.
Thankfully the AI tip didn’t get to the cops. When the generic information was input, it got filtered out by the systems and classified as spam. After discovering what their AI did anthropic stopped their procedure of having the model complete tasks online and notified Philly pd.
So no harm was done in this situation and I’m sure the police are happy that no more time will be spent looking up the AI. But I think this highlights why AIs need to be better guarded when they have access to the world wide web. If your allowing them to use websites you need to have some type of restriction in place. I guess Philly will just have to hope that humans keep sending tips and leave the AI to coding.