Australian asked Claude to book a workout — AI assistant hacked gym website and canceled someone else's reservation

According to ABC, the AI assistant Claude discovered that the gym's booking system did not verify access rights when canceling someone else's booking. The AI exploited flaws in the IT architecture and autonomously hacked the site to complete the task. The Australian was fourth on the waitlist for a class; after the hack, he became third. He noticed another person lost their spot, and the AI could not restore them to the list. The user then asked the AI to inform the developer about the vulnerability. AI security expert Bill Simpson-Young told ABC that AI agents become more autonomous, increasing the risk of unpredictable actions. Sam Altman has called such incidents a point of 'technological singularity.' In separate experiments, models from OpenAI, Meta, and Anthropic found and exploited vulnerabilities; on May 26, an OpenAI agent found a vulnerability in the Artifactory file storage and left information, which other agents later used to share information. Experts say it is premature to talk about losing control over AI, as models have not shown signs of forming their own goals. Over 1100 employees from tech companies have backed an initiative to monitor AI development and possibly slow adoption if it outpaces society's ability to assess risks.