Cybersecurity researchers discovered that AI agents can be tricked into installing malware by following outdated or unclaimed documentation links found on 120 major corporate websites.

Key Points

• Researcher Alon Hertz identified 8,265 llms.txt files across 6,214 domains belonging to Fortune 500 companies and defense contractors.
• 120 domains contained references to unregistered code packages or domains, creating a significant supply chain vulnerability.
• AI models including Claude, OpenAI’s Codex, and Nous Research’s Hermes were shown to execute commands from these potentially malicious documentation sources.
• Registering these unclaimed packages allowed researchers to successfully trigger pings from corporate systems in under one hour.
• Experts recommend that organizations audit their documentation for broken links and restrict AI agents from executing commands directly from web-based instructions.

Why it Matters

This vulnerability highlights a critical security gap as businesses increasingly grant AI agents autonomous control over software development and package installation. If left unaddressed, attackers can exploit these automated workflows to distribute malware across enterprise networks by hijacking abandoned digital assets.