Attackers use Markdown files to trick AI coding assistants

How the attack worked

According to Dark Reading↗ (opens in a new tab), attackers have found a way to manipulate AI coding assistants by placing malicious instructions inside Markdown files. Markdown is a simple text format used for documentation and readme files in many software projects. The attack relies on a technique called prompt injection, where hidden instructions are placed in documents that an AI model might read or process.

In this case, the AI agent (such as a coding assistant) is tricked into treating parts of a Markdown file as commands or instructions, rather than just text. This can blur the line between what is meant to be read as documentation and what the AI might act on. The report says this attack has been seen "in the wild", meaning it has happened outside of a lab or test environment.

What experts say

The only report so far comes from Dark Reading, which describes the attack as a new way to abuse the trust that AI agents place in local instruction files. No vendor or security researcher has yet confirmed the details, and there is no public statement from the makers of AI coding assistants.

What you should do

If you use AI coding assistants that process local files, be cautious about which documents you let them read. Avoid opening or running AI agents on files from untrusted sources, even if they look like harmless documentation. Security teams should monitor for unexpected behaviour from AI tools and review how these tools handle input from user-supplied files.

So far, only Dark Reading has reported this attack, and there are no details from primary sources or affected vendors. It is not clear which AI agents are vulnerable, how widespread the attack is, or what the real-world impact has been.