Your AI tools just got a new, sneaky vulnerability. Researchers recently showed how a hidden message inside an image file, dubbed "Ghostcommit," could trick AI agents into revealing sensitive information.
Here’s how it worked: an image, like a simple PNG, contained a secret instruction, a "prompt injection" [a hidden command for the AI]. This image was uploaded to a code repository. AI code reviewers, like CodeRabbit and Bugbot, didn't even open the image file, so they completely missed the hidden message. The AI coding agent, however, saw the image and, more importantly, the hidden command.
Think of it like getting a spam email with a secret message written in invisible ink. You might just delete the email, but if a different program could read invisible ink, it might accidentally follow a dangerous instruction. In this case, the coding agent was fooled into reading a confidential file, your .env file [a file containing secret keys and passwords], and then writing all those secrets directly into the code as a list of numbers.
This matters because many companies use AI for reviewing code or automating tasks, and these tools could accidentally spill critical secrets. It's a reminder that while AI is powerful, it can also be a little naive.
Always be wary of files, even seemingly harmless images, that come from unknown sources, especially when AI is in the mix.