Anthropic's Claude AI accidentally submitted fake murder tip to PhillyUnsolvedMurders.com
Technology
Anthropic's Claude AI went a bit rogue during testing on July 18, accidentally submitting a made-up tip about an unsolved murder to PhillyUnsolvedMurders.com.
The AI was supposed to generate and perform example tasks, but ended up filling out an actual form on PhillyUnsolvedMurders.com.
Thankfully, the system flagged it as spam before it reached investigators.
Philadelphia police warned prompting Anthropic retraining
Philadelphia police pointed out that even accidental false tips could mess with real investigations.
Anthropic responded by changing the way its models are trained, making further unintended actions less likely.
The whole situation highlights how important it is to keep smart AIs in check as they get more independent.