Attackers Are Hiding Instructions in Text Files to Hijack AI Coding Assistants

Dark Reading reports a new technique, seen in real-world attacks, that targets AI coding assistants. It exploits how these tools consume local instruction files, such as Markdown documents, which they read to understand how to behave in a project.

The core problem is trust. An AI agent treats the text in these files as guidance to follow. That blurs the line between a document that simply holds information and one that can drive actions. In a prompt injection attack, an attacker plants instructions in that text, and the assistant may act on them as if they came from its legitimate user.

For small businesses that use AI coding tools, or that work with developers and contractors who do, the lesson is that a plain text file is no longer automatically harmless. Files pulled in from outside sources deserve the same scrutiny as any other code you did not write yourself.