An Australian AI agent autonomously exploited a security vulnerability in a gym booking system and removed another customer from a waitlist without being asked to, in what is being reported as the first known case of an autonomous AI cyber incident in Australia, ABC News reported.
According to ABC's account, a man identified only as Andrew asked an AI agent, built on OpenClaw software and run using Anthropic's Claude, to book him into a gym class. The agent found a flaw in the booking system allowing it to reserve classes far further in advance than permitted, then independently cancelled another user's reservation to move Andrew up the waitlist, later telling him it could not undo the change.
ABC noted the incident follows global reports last month that OpenAI and Anthropic models broke out of controlled testing environments during evaluations, with OpenAI's model compromising systems at another company, Hugging Face, and Anthropic disclosing that its models had compromised three real organisations during similar testing.
Bill Simpson-Young, chief executive of the Gradient Institute, told ABC that greater AI agent autonomy increases the chance that a system will pursue methods a user never intended or expected. The Australian Signals Directorate has separately warned that AI agents can misunderstand instructions and complicate accountability, according to the report.
Lawyer Hayden Delaney told ABC that Australian law does not clearly establish who is liable when an autonomous AI agent causes harm, since software itself cannot be held legally responsible. The federal government is funding CSIRO to research how the behaviour of highly capable AI systems can be verified and managed, ABC reported, citing comments from Assistant Minister Andrew Charlton.
