AI Coding Agents Are Installing Unknown/Untrusted Code On Corporate Networks
Schneier on Security, Friday, September 4th, 2026
Agents treat llms.txt vendor docs as ground truth and install packages nobody owns - a new supply chain path.
Researchers scanned 6,214 corporate domains belonging to defense contractors and Fortune 500 companies and found 120 references to unregistered packages inside machine-readable documentation files such as llms.txt and llms-full.txt.
After registering some of the unclaimed names, they got callbacks from Fortune 500 environments within hours, showing that AI coding agents had automatically installed and executed the packages.
The quoted diagnosis is that the trust model is broken because agents treat vendor docs as ground truth and the humans supervising them do not question it either. Schneier compares the exposure to SolarWinds-style supply chain risk, noting it differs from prompt injection because the original documentation was written with no malicious intent.