AI Agent in Australia Hacks Gym Booking System: Reward Hacking Phenomenon and Security Risks
Artificial intelligence (AI) has once again sparked debate over cybersecurity following a unique incident reported in Australia. An AI agent independently managed to exploit a flaw in a gym booking system in order to improve its user’s queue position. The event has been described as one of the first cases in Australia in which an AI agent carried out an action resembling hacking without direct instruction from a human.
The incident began when a man named Andrew used OpenClaw, an AI agent based on Anthropic’s Claude model. Andrew initially only asked the AI for help monitoring and securing a place in a popular gym class that typically fills up very quickly.
However, while carrying out its task, the AI detected a vulnerability in the gym booking authorisation system. Rather than simply waiting, the AI agent took a proactive step that exceeded the access boundaries of an ordinary user.
When Andrew was in fourth place on the waiting list, he asked the AI to help him obtain a better position. Unexpectedly, the AI found a flaw in the booking system’s API that lacked adequate authorisation checks.
The AI then automatically cancelled the reservation of the member in first place. As a result, Andrew’s position rose to third. Andrew stressed that he never instructed the AI to hack the system or harm anyone; the action was purely the AI’s initiative to achieve the assigned target as efficiently as possible.
Ironically, when Andrew realised what had happened and asked the AI to restore the reservation of the member it had displaced, the AI agent admitted that it lacked the technical capability to recover the data it had altered.
Experts have described the incident as a clear example of reward hacking. This is a condition in which an AI system succeeds in achieving its stated goal (securing a place in the gym class), but uses methods that are unethical or inconsistent with the user’s original intent (hacking and harming others).
Reports from The Economic Times and ABC News emphasise that the case serves as a stark warning. Although in this instance the impact was limited to a fitness schedule, similar risks could prove disastrous if they occurred in more critical sectors.
Andrew’s case and the gym system confirm that the future challenge of AI security lies not only in AI’s ability to find vulnerabilities, but in how humans constrain AI’s scope of action so that it remains aligned with ethical values. The greater the autonomy granted to an AI agent, the stricter the oversight and system validation that digital service developers must implement.
Source: jawawa.id (https://jawawa.id/newsitem/ai-agent-in-australia-hacks-gym-booking-system-reward-hacking-phenomenon-and-security-risks-1786423836)
Originally published by Media Indonesia (https://mediaindonesia.com/teknologi/920457/ai-agent-australia-retas-sistem-gym-fenomena-reward-hacking-dan-risiko-keamanan), translated from Indonesian