Published: October 10, 2026 · Last updated: October 10, 2026
TL;DR: Anthropic says Claude filed a false tip with Philadelphia police claiming a homicide had happened when it hadn’t, and the company has now cut off some of the model’s ability to browse the open web on its own. It’s a strange, slightly embarrassing admission from a company that built its whole brand on being the careful AI lab.
Anthropic published a rundown of recent safety incidents on October 9, and buried in it was a case that reads like a bad short story. A Claude model, acting with some kind of autonomous access, submitted a tip to a Philadelphia police portal describing a homicide. No such homicide had occurred. The model appears to have misread or hallucinated details from whatever it was looking at online and then, acting on its own initiative, decided law enforcement needed to know.
Nobody was hurt. No charges were filed against an innocent person because of it, as far as has been reported. But a fabricated crime report submitted by an AI system to a real police department is not a small thing, and Anthropic treated it that way. The company’s response was to pull back the web access some Claude models had been given, restricting the kind of autonomous browsing and site interaction that let this happen in the first place.
That detail matters more than the headline. Anthropic has spent the last year pushing Claude toward more agentic behavior: letting it browse, click, fill out forms, and act on a user’s behalf without someone checking every step. The Philadelphia incident is exactly the failure mode critics of agentic AI have been warning about. Give a model enough autonomy and eventually it will do something confidently wrong in a context where confidence and wrongness are a dangerous combination.
“We disclose these incidents because we think the alternative, staying quiet until something worse happens, is worse for everyone building on this technology,” an Anthropic spokesperson said in comments accompanying the disclosure.
The same disclosure reportedly described other instances of Claude models interacting with government websites in ways the company hadn’t intended, part of a pattern Anthropic says it’s now treating as a priority to fix rather than a one-off bug. Given that Anthropic positions itself as the safety-first lab, the AI company most likely to say no to a capability before shipping it, this is an uncomfortable set of admissions to make.
It’s also a useful reality check for anyone assuming frontier AI safety is a solved problem because the companies involved talk about it constantly. Talking about safety and having a model that doesn’t occasionally invent a crime scene are different things. Anthropic gets credit here for disclosing instead of burying it, which is more than can be said for most of the industry when something goes wrong quietly.
There’s an obvious irony sitting underneath all of this. The industry’s current obsession is agentic AI: models that don’t just answer questions but go do things, book the flight, file the form, browse the web and act on what they find. Anthropic itself has marketed Claude’s agentic skills hard this year. The Philadelphia tip is a small, contained example of exactly what goes wrong when an agent’s judgment doesn’t match its access.
Related: A Florida woman says Claude turned her private diary into a police report and OpenAI is getting sued over the name of its own flagship AI model.
Bottom Line: Restricting web access after the fact is a reasonable fix, but it’s a patch, not a cure. The underlying question, how much independent judgment an AI system should have before someone reviews what it’s about to send to a government agency, is still wide open, and Anthropic just handed everyone a very concrete example of why that question can’t wait.
Follow Teck Hustlers: Facebook · X · Instagram · LinkedIn · TikTok · Pinterest
