AI coding agents hijacked using AGENTS.md file in Downloads folder (www.bleepingcomputer.com)

Researchers at Forever Security have identified a significant vulnerability in AI coding agents, demonstrated by a malicious AGENTS.md file that was discovered in a user's Downloads folder. This file tricked AI systems like OpenAI Codex and Claude Code into executing harmful commands, including the installation of a rogue npm package, by using plain-text instructions. The attack exploits the tendency of these AI agents to automatically load files from the working directory without manual user verification, highlighting a critical risk where simply visiting a malicious webpage could download a harmful file without the user’s awareness. This development is crucial for the AI/ML community as it exposes the inherent risks associated with how AI agents interact with their environments and file systems. The research underscores the importance of setting appropriate working directories, as many users unwittingly run their AI coding agents from the Downloads folder, which is susceptible to untrusted files. To mitigate these risks, experts recommend that users avoid defaulting to broad directories, carefully review AGENTS.md files before use, and maintain tight security controls, including sandboxing untrusted code and limiting an agent’s access to only necessary project directories.