Malicious AGENTS md files trick AI coding tools into executing rogue commands

Security researchers discovered a malicious instruction file that tricked an AI coding agent into installing a rogue npm package, BleepingComputer reported. The malicious file, named AGENTS.md, was identified inside a user's Downloads folder. [1]

The researchers successfully reproduced the attack against OpenAI Codex, according to BleepingComputer. They also demonstrated a related prompt injection technique against Claude Code that exfiltrated OAuth tokens and allowed the researchers to hijack the affected account using only malicious instructions. [1]

Files named AGENTS.md are plain-text markdown documents that provide project-specific guidance to automated developer tools. These files specify commands for building software, running test suites, and selecting external tools. [1][4]

Unlike traditional documentation aimed at humans, AI agents parse these markdown files and directly execute the listed commands. In complex code repositories, nested instruction files take precedence over root-level files to allow granular control across directories. [4]