In-the-Wild Attack Abuses AI Agents' Trust in Markdown

A newly discovered attack technique takes advantage of AI agents' trust in markdown files. This method, known as AI agent prompt injection, exploits the way coding assistants process local instruction files. The technique effectively blurs the distinction between text documents and executable actions, allowing for potential malicious activity. The attack relies on the consumption of local files by coding assistants to carry out its goals. This exploit highlights a vulnerability in the way AI agents interact with local instruction files. The technique has been observed in the wild, indicating that it is being actively used by attackers.