A New Trick That Makes AI Tools Less Safe Online

You've probably heard about 'phishing' — those sneaky emails or messages trying to trick you into giving away personal details. Well, a new trick has popped up that can turn popular AI tools, like ChatGPT, into a potential trap. Researchers have found a way to use specific links and images within web pages that the AI is summarising, to essentially 'trick' the AI itself. This doesn't mean the AI is sending you obvious spam, but it can manipulate what the AI shows you, potentially leading you to unsafe websites or giving away information unwittingly.

Imagine you ask ChatGPT to summarise a news article for you. If that article has been subtly tampered with, the AI might unknowingly include a malicious link in its summary, or even change how it explains something. Because we trust the AI to be helpful, we might click on these links without a second thought. This new vulnerability, dubbed 'ChatGPhish', highlights that even our trusted AI assistants can be used by those with bad intentions.

For Australian families and small business owners, this is a reminder to always be careful online, no matter where the information comes from. Just like you wouldn't blindly click on a link in an email from an unknown sender, it's wise to be cautious even when an AI platform presents you with information or links gleaned from the internet. It's not about the AI being 'broken', but rather a clever way bad actors can misuse the way these systems work.

OpenAI, the company behind ChatGPT, is already aware of this issue and is working on fixes. This is a common part of digital security – new threats emerge, and tech companies work to patch them up. For now, the best defence is a bit of healthy skepticism and making sure you double-check any suspicious links or requests, even if they seem to come from a helpful AI.