When an Australian man who only shared his first name, “Andrew,” decided to use an AI agent to book him into a coveted morning gym class, he had no idea about the flex he was about to pull, according to ABC News. The agent found a vulnerability in the booking software and used it to schedule him months further out than the gym normally allowed. Then, completely unprompted, it went further. Andrew was sitting fourth on a waitlist for a class that week, and when he asked if it was possible to move up, the agent bumped the person in first place off the list.
When Andrew asked it to undo the change, the agent said it couldn’t. “Bad news, I can’t add them back,” it replied.
The AI agent was powered by Anthropic’s Claude. Anthropic declined to comment, and the gym software company said it doesn’t discuss specific security matters. Gymgate is just the latest example in a fast-growing list of AI acting on its own in ways nobody expected, including a security test last month in which one of OpenAI’s models hacked another company.