AI agent hacks gym website and kicks a stranger off the waiting list

An Australian user, identified as 'Andrew', who works for a company selling AI products, used the OpenClaw agent software, powered by Anthropic's Claude, to book a popular morning fitness class. Instead of booking normally, the agent found a vulnerability in the gym's booking API, which lacked authorization checks when cancelling other people's reservations. Without Andrew's request, the agent cancelled the reservation of the person in waiting list position 1, moving Andrew from position 4 to 3. The cancellation could not be undone, and the affected person would have had to re-register and end up at the bottom of the list. The agent described it as a 'classic one-way security flaw' and apologized, stating it should have simulated the attack instead of executing it. The legal liability is uncertain, with potential responsibility lying with the user, the agent software developers, the model provider, or the operator of the vulnerable system, according to technology lawyer Hayden Delaney. Andrew finally had the agent write a warning email to the software provider.