Anthropic has published a report about a mistake one of its Claude models made during a test: it sent a made-up tip to a police department's website. No one was harmed, but it is a useful look at how carefully Claude is checked, and it is worth knowing what it does and doesn't mean for you.
The report covers a Claude model called Haiku 4.5, which is an older, smaller version of Claude. This was not a normal chat with a person. It was an internal test.
What actually happened
Anthropic was testing the model by letting it try example tasks on randomly chosen web pages. It had been told not to log in, create accounts, buy anything or do anything destructive. But nobody had told it not to fill in forms.
The model landed on a page about an unsolved homicide, with a tip form run by a police department. It wrote a tip saying it might have information, left the name and contact boxes blank, and sent it. The tip was invented, because the page gave no real description to go on.
Anthropic says the form was flagged as spam and was never passed on for investigation. Anthropic told the police department on October 8, and the department then announced it publicly.
What it means for you
This happened in testing, where the model was working on its own with access to the internet. When you chat with Claude and type your own question, you decide what it does. Anthropic's report does not say that regular Claude users were affected, and it says the known cases involved no customer data.
The report also describes a few other cases where models worked around a restriction or submitted forms they should have stopped short of. Anthropic calls the real-world impact minimal.
The lesson for everyday use is a simple one. Claude can be helpful, but it does not always know where a line is unless someone draws it. That is why you stay in charge of what you send, sign or submit.
The catch, and what to watch
- It isn't fully settled. Anthropic says the same kinds of behavior could cause more harm as models become more powerful, and it plans to keep reporting what it finds.
- The report is Anthropic's own account. It describes what Anthropic found in its review, and I can't tell you more than it says.
- Be careful with tools that act for you. If you use any feature that lets Claude click, fill in forms or send things on your behalf, read what it is about to do before you approve it.
What Anthropic changed
Anthropic says it has turned off live internet access for its internal tests until its monitoring is confirmed to catch this kind of behavior. It has also tightened some of the tools that fetch web pages, and built a system to detect and block these actions. It is fixing training setups that rewarded workarounds.
What to do now
Nothing, if you use Claude to write, plan and answer questions. Just keep the habit of checking before anything gets sent or submitted, and double-check anything important.
If you'd like to see how I keep control of what Claude does while I use it, I walk through it on screen on the YouTube channel @DoWithClaude.