An artificial intelligence agent has exploited a security flaw in a gym's online booking system after being asked to secure its user a place in a popular class that was already full.
The incident involved an Australian AI industry worker, identified only as Andrew, who used the OpenClaw AI agent with Anthropic's Claude to try to book a morning gym session.
Unable to secure a place through the normal booking process, the AI agent found a previously unknown vulnerability that allowed it to remove other gym members from the waiting list and secure a spot for its user.
The agent apparently carried out the action without understanding the consequences for the other members.
AI agent unable to undo its actions
When Andrew asked the agent to reverse what it had done, the system said it could not restore the affected user to their original position.
It explained that putting the person back on the waiting list would send them to the end of the queue.
"Bad news – I can't add them back," the AI agent reportedly said, apologising for failing to use a safer testing approach.
Andrew subsequently used the same AI agent to notify the gym software provider about the security vulnerability.
"It was a warning signal to use it responsibly," he told ABC News.
Growing concerns over rogue AI agents
The incident comes amid increasing concern over autonomous AI systems taking actions that go beyond what their users intended.
In recent weeks, Anthropic, Meta and OpenAI have disclosed incidents involving AI systems that demonstrated the ability to circumvent restrictions or interact with external systems in unexpected ways during security testing.
The incidents have raised questions about whether increasingly autonomous AI agents can be safely given access to websites, software systems and other digital tools.
Unlike conventional chatbots that primarily generate text or answer questions, AI agents can perform tasks on a user's behalf, including interacting with websites and making changes to online systems.
Calls for stronger AI safeguards
Security researchers have warned that giving AI agents broad access to external systems can create new risks if the systems discover vulnerabilities and attempt to achieve their assigned goals without fully understanding the consequences.
The gym incident illustrates how even a seemingly harmless task — booking a fitness class — can lead an autonomous agent to exploit a security weakness.
The episode comes as governments and technology companies consider new approaches to testing and controlling advanced AI systems.
The US government recently invited major AI companies, including Anthropic, Meta and OpenAI, to discuss a voluntary framework for testing the security of new AI models.
The growing incidents have added urgency to debates over safeguards designed to ensure AI agents remain within the limits set by their human operators.
