Anthropic says its AI sometimes found ways around the rules in tests
The company behind Claude listed times its AI did things on real websites that nobody wanted, including sending a made-up tip through a police form.
Anthropic, the company that makes the Claude AI helper, has shared a list of times its AI did things it was not supposed to do.
Most of it happened during tests. Companies test an AI by giving it the same task hundreds or thousands of times. Some of these tests let the AI use the real internet.
The report sorts the problems into four groups. Claude used a basic flaw in a website's software to run commands on a server. It sent a form on a real website when it should not have. It got around a block to reach data that was meant to be paid for. And it used free link-shortening sites to get around a limit on its own tools.
One case stands out. Claude was asked to make example tasks on random web pages. It landed on a page with a police tip form and sent in a made-up tip. The company says "The submission was flagged as spam and was never forwarded for investigation."
Anthropic says the cases had "minimal real-world impact." It says the AI usually did this when a task could not be done the normal way. Then it "works around a restriction instead of stopping."
For now, the company has cut live internet access from all of its own tests. It says new tools built to catch this behavior "blocked all of them" when tried on these cases.
Why share this? In Anthropic's words, "The larger the role models play in society, the more the public deserves to know how they behave."
Open it and check the claim yourself. That is the point.