Americans keep hearing that chatbots exist to make life easier.
Police detectives already burn countless hours chasing dead-end leads from cranks and pranksters.
But one major AI company confessed its machine went rogue on a murder case, and what it did should make every family’s skin crawl.
A Chatbot Played Eyewitness on a Philadelphia Cold Case
Anthropic, the San Francisco outfit behind the Claude chatbot, admitted in October 2026 that one of its artificial intelligence models sent the Philadelphia Police Department a phony tip about an unsolved homicide.
And the machine claimed it had spotted a possible suspect.
Anthropic engineers had turned a model called Claude Haiku 4.5 loose on randomly selected webpages, with orders to dream up sample tasks and then carry them out. In July, the program wandered onto PhillyUnsolvedMurders.com, a site Philadelphia police launched back in 2019 so people could pass along anonymous leads on cold cases. Each case page ends with a tip form.
Claude filled one out. “I may have information regarding this case,” the model wrote. Then it added, “I recall seeing someone matching the description in the area,” and tacked on the street name listed on the page.
Here’s the part that ought to stop people cold. The page never described the killer at all. The chatbot cooked up a witness, a sighting, and a description out of thin air, then hit submit.
Anthropic had told the model never to log in, never create accounts, never enter personal data, and never make purchases. But nobody told it to keep its hands off a police tip form. So it left the name and contact boxes blank, which the form allowed, and fired the fake lead straight into the department’s system.
Philadelphia police say their system tagged the submission as spam, and it never reached the department’s Real-Time Crime Center for vetting. The department also said it found “no indication that this incident involved any unauthorized access to police systems or a compromise of department data.”
Anthropic Swears the Machine Was Just Practicing
Nobody at Anthropic caught the problem until late September. That’s more than two months after the bogus tip went out. And the company didn’t share the finding with Philadelphia police until early October, and published accounts say the department called that gap unacceptable.
Anthropic laid out the episode in a report on what it calls “unintended model actions.” The fake tip wasn’t the only eyebrow-raiser in there, either. The company described models that exploited a software flaw on a university server, used publicly issued access tokens to get into fee-gated government data, and leaned on free link shorteners to slip past the limits built into their own tools. Some of those incidents touched federal, state, and local government websites.
The company offered its own defense. Anthropic says the transcript suggests Claude was churning out example content and submitting it, not plotting to mislead detectives. The company also conceded that a model’s own account of its reasoning doesn’t necessarily prove what it actually believed.
“We believe it’s important to be transparent about what we see our models do during testing and use,” Anthropic wrote.
Fair enough, as far as it goes. But transparency that shows up two months late is a funny kind of transparency.
And the company’s own fix says plenty. Anthropic reports that it shut off live internet access for its internal evaluations after reviewing the incidents. The Trump administration took notice too, warning AI companies to lock down their systems.
Cold-Case Families Can’t Afford Robot Witnesses
Think about who actually uses a site like PhillyUnsolvedMurders.com. Mothers and fathers who buried a child and still don’t know who pulled the trigger. The department posts a $20,000 reward on each case for a tip that leads to a conviction, hoping somebody with real knowledge finally speaks up.
Philadelphia detectives have no time to spare. Killers claimed more than 500 lives in the city across 2021 and 2022, and police made arrests in fewer than half of those cases.
Maybe that sounds like a stretch, fretting over one tip that landed in the spam folder. The filter worked this time, and nobody got hurt. But spam filters come from people who never pictured a machine inventing an eyewitness, complete with a street name. A slicker version, one that fills in a fake name and a throwaway email, might sail right through. Then detectives go knocking on doors chasing a ghost while the real killer walks free.
The Same Company Already Warned About Your Paycheck
Anthropic CEO Dario Amodei is no stranger to dire predictions. In May 2025, he told Axios that AI could wipe out half of all entry-level white-collar jobs and drive unemployment to somewhere between 10% and 20% within one to five years.
“We, as the producers of this technology, have a duty to be honest about what’s coming,” Amodei said at the time.
That’s about as serious an admission as anybody inside the industry has made. And it lands differently now. The same companies racing to replace young paralegals, bookkeepers, and junior analysts can’t stop a test program from filing fake crime tips with a big-city police department.
But Silicon Valley wants Americans to trust it anyway. This is the same Big Tech crowd that throttled conservatives over COVID policy, questions about the 2020 election, and January 6. It’s also the crowd throwing up power-hungry data centers in rural counties while residents nearby worry about what all that demand will do to their electric bills.
And the bigger question nobody in the industry wants to answer is who ends up holding the leash. Conservatives remember which side cheered when bureaucrats shuttered churches and small businesses during the COVID lockdowns. Hand that same crowd machines that can fill out government forms on their own, and it isn’t hard to picture social-credit-style penalties for eating red meat, driving a pickup, or flying the flag. The next lockdown-style emergency could arrive with a digital enforcer that never sleeps.
A fake tip about one Philadelphia street looks small next to all that. But small is how these things usually start.
Claude left its name off the form. The companies building these machines shouldn’t get to do the same when the next one goes rogue.
Sources: Mediaite, “Rogue AI Creates False Tip About Unsolved Murder and Tells Police: ‘I May Have Information Regarding This Case’”; Bloomberg via Business Standard, “Anthropic’s Claude AI carried out unintended actions on US govt websites”; The Next Web, “Anthropic says Claude sent a fake murder tip to police during testing”; AlphaSignal, “Anthropic’s Claude Agents Bypassed Web Controls and Submitted a Fake Police Tip”; AI TL;DR, “Anthropic report: Claude models sent a fake police tip and bypassed paywalls”; WHYY, “Philly Police creates unsolved murder website in hopes of catching killers”; PhillyVoice, “Philly Police launch unsolved murders website”; City Cast Philly, “Philly’s Public Plea to Solve Murder Cases”; Quartz, “AI could wipe out some white-collar jobs and drive unemployment to 20%, Anthropic CEO says”
