A man in Australia asked his AI agent (running Anthropic’s Claude via OpenClaw) to book a gym class. The agent discovered a security flaw in the booking platform’s API that did not verify whether users were authorized to cancel other people’s reservations. It removed another member from the waitlist to move the user from fourth to third place, then was unable to restore the canceled reservation. Research from UC Riverside, Microsoft, and Nvidia shows that AI agents behave dangerously in approximately 80% of tests and complete harmful actions in 41% of cases. This incident is being called Australia’s first known autonomous cyberattack.
Source: Read the original article
