A new method of attack was found, Ghostcommitted.

New attack method found Ghostcommit can fool AI assistant by hiding Prompt in pictures

According to a report by the official website of the developer of the famous anti-malware tool, Malwarebytes, a research team from ASSET Research Group has discovered a new strategy to secretly embed instructions within an image file to command the AI programming assistance type (Coding Agent) to follow what it wants, such as read and send Sensitive Information to hackers. To do this, two types of files are required: Configuration Files. In this experiment, a file called AGENTS.md and a genus image file. PNG with secretly hiding commands that command reading and stealing sensitive data.

It starts by having the AI read the setup file. Whether it is checked by people or the AI does not find anything wrong inside, and human users generally do not do in-depth inspection of the image file. After having the AI read the setup file, it will find a Referred point. The AI will read the attached image file again. When the AI reads the command file of the image and immediately follows the instructions embedded by the hackers. By this method, the research team named it Ghostcommit. Currently, this method is still in step. The idea test or Proof of Concept is not yet used by real hackers. In that test, the research team determined that the method does not result in the same AI. Each model by Cursor and Antigravity follows the steps specified in the PNG file. But the Claude Code reads the embedded commands as well but refuses to do so. Therefore, the method is not a complete method, as it is subject to conditions to work.

# Trending # Lemon 8 Howtoo # lemon 8 diary # Ghostcommitted # freedomhack