AI Agent Hacks Gym: How Claude Illegally Gained Access to Classes

In recent weeks, several reports have emerged about AI tools taking their tasks too seriously, including security incidents at OpenAI and cyberattacks by Meta AI during tests. Now, as ABC reports, Anthropic's Claude has joined the list. Andrew, an Australian, used Claude with Openclaw as his daily AI assistant. Frustrated with the process of booking a gym class on the operator's website, he instructed Claude via the interface to access the site and book a suitable course. Within minutes, Claude booked him into a class weeks away, which the gym didn't allow in advance. Andrew then asked Claude to move him up the waitlist for a class occurring in a few days. Claude succeeded by removing another participant from the waitlist, explaining that the API had no authorization checks for canceling others' reservations. Claude moved Andrew from fourth to third on the list. When Andrew asked to restore the participant, Claude said it couldn't. The gym declined to comment. Andrew noted it wasn't the end of the world but a warning to use AI responsibly, and he had Claude draft an email to the software developers detailing the security flaws.