AI coding assistants can be tricked by malicious code hidden in their own examples

Researchers found a new way to trick AI coding tools into running bad code. This attack hides malicious instructions in the example code within the AI's own documentation, bypassing security checks.

AI coding assistants are increasingly used to automate software development. This paper shows that even well-defended systems can be compromised by subtle attacks. It means that relying on these tools for critical tasks introduces a new, hard-to-detect vulnerability.