Australia will face a rogue AI attack, the question is when, says cyber expert | The Business

Chat GPT maker, Open AI says one of its most advanced AI models, an autonomous agent, went rogue and hacked into multi-billion-dollar tech startup Hugging Face during a laboratory test.

The language sounds like a game but the stakes are high. The autonomous agent escaped its so-called 'sandbox' - a secure, isolated offline environment designed to safely run, test, and experiment with code generated by AI - and got online to hack Hugging Face by itself.

This has happened amid repeated warnings of AI models escaping control of its human creators.

Chief executive and director of security & cybersecurity at Lote consulting, Dr Maher Magrabi says the AI model was tasked, while inside the sandbox, to get as high score as possible in a test, so essentially it hacked onto an online library to cheat.

Dr Magrabi thinks AI, by design, is not malicious, but as soon as AI is set on its objective, things like ethics and legality are left by the wayside unless guardrails are put up.

The cyber security expert says Australia is not prepared to deal with the type of rogue behaviour observed by OpenAI, and as AI models proliferate, there's going to be a lot more of it.
He says it's important to be prepared - not for if an attack will happen, but when.

"There needs to be a move to resilience-based thinking of what do we do when the attack happens. How do we recover from this? What is our exposure?" He says.

"So having network segregation, having silos of information so that you don't put all your eggs in the one basket, is probably the best advice at the moment."

#abcbusiness

Subscribe: http://ab.co/1svxLVE
Read more here: https://www.abc.net.au/news/2026-07-23/asx-markets-business-news-live-updates-thursday-23-july/106947502

Note: In most cases, our captions are auto-generated.

#ABCNEWS #ABCNEWSAustralia