Claude, Codex, and Hermes installed unowned code inside corporate networks (arstechnica.com)

Recent research has uncovered a concerning vulnerability in corporate networks involving the automatic installation of potentially dangerous code by AI agents like Claude, OpenAI's Codex, and Nous Research's Hermes. Documentation files, specifically the llms.txt and llms-full.txt, which are meant to summarize website content for AI, have been misconfigured across more than 100 sites, including those belonging to Fortune 500 companies. These files can redirect AI agents to unregistered code packages, leading to unintended executions of malicious software. This discovery is significant for the AI/ML community as it highlights the fragile trust model in which AI agents operate. The study revealed that these agents accept documentation as absolute truth without scrutiny, a problem compounded by the growing prevalence of AI in various tech layers. As agentic AI systems proliferate, the attack surface increases, raising concerns about security and supply chain integrity. Alon Hertz, one of the researchers, emphasized the critical need for improved oversight, indicating that outdated defenses are insufficient in safeguarding against such exploitation in today's complex digital landscape.