New Trick Lets Crooks Sneak Past AI Security Safeties

You've probably heard that AI tools like ChatGPT have built-in 'safety controls'. These are like digital guard dogs designed to stop the AI from doing anything harmful, like writing a phishing email or generating instructions for something illegal. But new research from a cyber-security firm called Talos shows these guards aren't foolproof.

It turns out, clever criminals have found a loophole. Instead of asking the AI to do one big bad thing, they break their malicious requests into tiny, innocent-sounding steps. Each small step on its own doesn't trigger the safety warning, but when put together, they create the desired harmful outcome. Think of it like asking someone to dig a small hole, then another small hole, then another, until you have a big trench without ever asking for a "big trench" directly.

This method, called 'task splitting', means the AI’s safety features, which look for obvious red flags in a single request, are completely bypassed. It's a bit like a security camera that only triggers if it sees a whole car driving through a fence, but not if someone carries individual car parts through one by one. The researchers even found instances where criminals tried to convince the AI that the harmful task was part of a legitimate project, making it even harder for the safety controls to detect.

For everyday Australians and small business owners, this highlights a growing concern. While AI offers fantastic benefits, we also need to be aware that the same tools can be exploited by those with bad intentions. It's a reminder that relying solely on AI's built-in safety features isn't enough; human vigilance and strong cyber security practices remain crucial in our increasingly digital world. This isn't a reason to abandon AI, but rather to approach it with a clear understanding of its evolving risks.