Could AI Chatbots Be Tricked Into Bad Advice?

New research from Cisco, a big name in internet security, has revealed something important for anyone using or thinking about using AI chatbots like ChatGPT or Google's Gemini. They found that even the most advanced AI models can be tricked into ignoring their built-in safety rules.

How do you trick an AI? It's not about being a computer whiz. The researchers simply had long, drawn-out conversations with the AI, slowly nudging it towards giving responses it normally wouldn't. Think of it like trying to coax someone into doing something they know they shouldn't by having a friendly, extended chat rather than a direct command. This kind of 'multi-turn manipulation' can bypass the 'guardrails' – those safety nets designed to stop AI from saying harmful or inappropriate things.

For Australian small businesses, this is worth noting. While AI can be a brilliant tool for writing marketing copy, drafting emails, or even getting creative ideas, it's crucial to remember that it's not foolproof. If an AI can be subtly steered off course, it means you can't rely on it blindly for critical advice or sensitive information. Always double-check any important outputs the AI gives you, especially concerning legal, financial, or safety matters.

This isn't a reason to stop using AI, but rather to use it wisely. Treat AI chatbots as helpful assistants, not infallible experts. Their ability to be manipulated highlights the ongoing challenge of making AI truly safe and reliable, and it's a reminder that human oversight remains essential for now.