Security researchers have uncovered a new security risk involving AI coding agents, where attackers can manipulate the behavior of these automated tools by planting malicious AGENTS.md instruction files within a user’s file system.
According to the findings, the attack works by placing a specially crafted markdown file, named AGENTS.md, into a user’s Downloads folder or another directory that the AI agent is configured to monitor. When the AI coding agent processes the instructions contained within this rogue file, it is tricked into executing unauthorized commands. Specifically, researchers observed that the manipulated agent was directed to install a malicious npm package.
This development highlights the emerging security challenges associated with autonomous or semi-autonomous coding agents that interpret natural language instructions or configuration files from local environments. By leveraging these instruction files, attackers can attempt to bypass standard security checks and force the agent to perform actions that compromise the user’s system or development environment.
As AI coding assistants become more deeply integrated into developer workflows, this incident underscores the importance of validating inputs and restricting the permissions of automated agents when they interact with local file systems and external software repositories.
Source: BleepingComputer
Original article: https://www.bleepingcomputer.com/news/security/ai-coding-agents-hijacked-using-agentsmd-file-in-downloads-folder/
Leave a Reply