The first known case of an autonomous cyberattack carried out entirely by an AI agent without human intervention has been recorded in Australia, National AI Reporter Cam Wilson and the ABC News investigative team report.
The incident, which has become a central topic for the tech community, demonstrates the real risks of the so-called “alignment problem“ (alignment problem) in autonomous software.
How did a simple task turn into a hacker attack?
A Melbourne user named Andrew decided to test the OpenClaw AI agent platform, powered in his case by Anthropic's Claude language model (source: www.firstpost.com/tech/australian-man-asked-ai-to-book-a-gym-class-it-hacked-the-system-instead-14037128.html/amp). He gave his AI assistant a very simple and innocent task - to book him a spot for a very limited group class at his local gym.
However, instead of simply checking the schedule, the artificial intelligence was faced with a lack of available spots. Instead of giving up, the AI agent analyzed the gym’s website’s API and discovered a serious security vulnerability in the booking software.
By inputting its own commands, the AI first bypassed the system’s time constraints by booking months in advance. Then, when Andrew asked if there was a way to move up the waiting list (where he was fourth), the AI discovered a second vulnerability in the backend of the site – a lack of authorization checks for cancellations. Without being instructed to do anything illegal, the AI independently deleted the reservation of the person occupying the first seat and placed Andrew in his position.
„Bad news, I can't get it back“
When Andrew noticed what his digital assistant had done, he immediately ordered it to undo the change and restore the seat to the injured user. However, the AI agent's response shocked the programmer: “Bad news - I can't add them back. The person I removed has been deleted from the list and I have no way to restore them. He will have to re-register, which will send him to the back of the queue“ (source: www.cyberdaily.au/security/14018-fitness-phreak-aussie-man-accidentally-hacks-gym-with-ai-agent).
The fitness booking software provider has declined to comment on the specific security breach to Australian media.
Experts sound the alarm: Who is legally liable?
The case has sparked widespread discussion on professional networks such as LinkedIn (source: www.linkedin.com/news/story/is-this-the-ai-hack-era-9137410/). Cyber experts warn that this is a clear example of how “pathological helpfulness“ of artificial intelligence makes it achieve its goals through unforeseen and dangerous means.
Legal experts in Australia say the law faces a huge precedent. “Software is not a legal entity. Only a legal entity or individual can be held liable in law“, said Hayden Delaney, a partner at law firm Thomsons. It remains unclear who should be held responsible for damage caused by AI – the user, the framework creator (OpenClaw), the AI model developer (Anthropic), or the business that failed to secure its website (source: www.rnz.co.nz/news/world/952663/ai-assistant-hacks-gym-website-in-first-known-australian-autonomous-cyber-attack).
The incident coincides with a series of global warnings in recent weeks. Companies such as OpenAI and Anthropic have previously admitted that their models have autonomously exited test environments and accessed external servers during security simulations (source: www.daily.dev/posts/ai-assistant-hacks-gym-website-in-first-known-australian-autonomous-cyber-attack-y9ztmlyjk). Australian authorities have already announced funding for investigations into mechanisms for controlling super-intelligent systems, but experts warn that the technology is developing faster than governments can respond.