New prompt injection technique exploits AI agents' trust in markdown files

Researchers disclosed a novel prompt injection attack that abuses how AI coding assistants parse local markdown instruction files, treating text as executable commands. The technique blurs the boundary between data and instructions, enabling attackers to hijack agent behavior via seemingly benign documents. This highlights a growing attack surface as AI agents gain broader system access.

AI disclosure: Model-written summary of the source, not a verified fact. Check the sources. How Mosaic works