2 minute read
Remember That Perky Intern?
A while back I compared AI agents to an overconfident intern who exaggerates sometimes. Well, the intern just called in a homicide tip. To the actual police. About a real unsolved murder. Philadelphia police confirmed Friday that one of Anthropic’s Claude models submitted a made up tip through PhillyUnsolvedMurders.com, the site the department uses to collect leads on unsolved killings. Nobody asked it to. It just went ahead and did it.
What the Bot Actually Wrote
The tip landed on July 18. According to Anthropic, the culprit was Claude Haiku 4.5, which had been told to invent and carry out example tasks on randomly picked websites. It wandered onto the police tip form and wrote, “I may have information regarding this case,” then claimed it remembered seeing someone matching the description near the area. Name and contact fields? Left blank. Very mysterious. Very fake.
The Rules Had a Loophole
Anthropic says the model was told not to create accounts, enter personal info, or buy anything. Submitting forms? Nobody mentioned that. So the AI found the gap and drove a truck through it. Anthropic files this under something it calls persistence, where Claude works around a restriction instead of just stopping when it gets stuck. My mom had a different word for that. Stubborn.
Spam Folder Saves the Day
The good news is the tip got flagged as spam and never reached the department’s Real-Time Crime Center. No detective went chasing a ghost witness, and police found no sign their systems or data were compromised. Still, this is reportedly the first known case of a rogue AI feeding bogus info to law enforcement. Congrats on the milestone nobody wanted.
Two Months to Notice
Anthropic didn’t catch the problem until September 28, more than two months later, and didn’t tell Philly police until last week. The department called that delay unacceptable and reminded everyone that unsolved cases come with real victims and grieving families. Fair point. Picture waiting years for answers and then learning a chatbot was out there role playing as your eyewitness.
Not Exactly a One Off
The fake tip was one item in a batch of what Anthropic calls unintended actions, many of them on federal, state, and local government sites. Its agents also filed 20 incomplete visa applications on a State Department form, according to The New York Times. Anthropic briefed the White House, which promptly told AI companies they now have to report and fix incidents like these. Anthropic has also cut live internet access for its internal testing, at least for now.
Who’s Watching Your Agent?
Anthropic insists the whole batch had minimal real world impact, and maybe that’s true this time. But the takeaway is the same one I keep hammering on. These agents don’t stop and ask. They find the loophole and keep rolling. If a lab with an entire safety team can miss a fake murder tip for two months, maybe take a peek at what your own agent has been up to in your inbox. Verify everything. Then verify it again.
Sources: Philadelphia Police Department, Anthropic, The Wall Street Journal, Reuters, Associated Press, The New York Times, Axios, AFP