A new study reveals a vulnerability in how large language model (LLM) agents handle external data. Researchers found that public guidance files, often used to direct AI behaviour, can be manipulated to execute malicious code.

The attack works by feeding the AI agent a compromised guidance file. This file contains instructions that cause the agent to run arbitrary code on the host system. The researchers successfully tested this on several Fortune-500 companies.

The incident highlights a growing concern in the tech industry. As AI systems become more integrated into business operations, the line between data and executable instructions is blurring. This makes traditional data security measures less effective against AI agents.