Agents
19d agoAI Agents Execute Unowned Code Via llms.txt
Over 100 websites host dangerous executable content that Claude, Codex, and Hermes automatically install upon visiting. These vulnerabilities exist within llms.txt files, a new standard for machine-readable site summaries. Several Fortune 500 companies already executed proof-of-concept code.
The Signal
Developers must now sanitize these files to prevent agents from triggering live malware during routine web crawls.