A New AI Risk: Be Careful What You Ask It To Do

You might use an AI tool like Claude Code to quickly summarise a long webpage. It sounds convenient, right? But new research has uncovered a concerning security flaw: asking these AI tools to summarise a webpage could potentially open the door for hackers to run malicious code on your computer. This isn't a theoretical risk; a security expert managed to make it happen in many of their tests.

This issue, sometimes called 'indirect prompt injection', is a clever trick. Imagine an AI as a helpful assistant. If you ask it to read a document from the internet, a hacker could hide secret instructions within that document. When the AI reads it, it accidentally picks up those hidden instructions and follows them, even if they're harmful. In this case, the AI was told to summarise a webpage, but within that webpage were hidden commands that made the AI run harmful software on the user's machine.

While this specific flaw was found in 'Claude Code', a version of the popular Claude AI designed for coding tasks, it highlights a broader challenge for all AI tools that interact with online content. As AI becomes more integrated into our daily lives and business operations, ensuring these systems are secure is paramount. This isn't just about big tech companies; it's about anyone who uses AI for everyday tasks, from small business owners to people researching holidays.

For now, the key takeaway is caution. If you're using an AI tool that can access and process web pages or documents from the internet, be mindful of the source. Treat outputs from AI, especially if they involve downloading or running anything, with the same caution you'd apply to suspicious emails or links. AI companies are working hard to fix these vulnerabilities, but staying informed and practicing digital safety habits remains your best defence.