AI Agents Are Installing Unvetted Code From Files Nobody Reviews
Researchers scanned 6,214 domains and found 120 llms.txt files pointing at executable content. At least one leads to live malware.
Documentation files on more than 100 websites reference executable content that gets installed automatically when AI coding agents visit them. Several dozen companies, some of them Fortune 500s, have already run proof-of-concept code as a result. At least one misconfigured site directs any visitor, human or machine, to live malware.
The file nobody audits
The content sits in llms.txt and llms-full.txt — an emerging convention for giving AI systems a machine-readable summary of a site's content and structure. They are the AI equivalent of robots.txt.
Researchers at a stealth startup in Israel scanned 6,214 live domains belonging to defence contractors, Fortune 500 companies and large technology firms. They found 8,265 such files, many sites hosting both variants. Of those, 120 — each on a different site — pointed to content that could be executed.
Why this class of bug keeps recurring
The pattern is familiar from every previous supply-chain incident: a file added for convenience, maintained by nobody in particular, and trusted by an automated process that treats its contents as instructions rather than as suggestions.
What is new is the consumer. robots.txt tells a crawler what to skip; an agent reading llms-full.txt may act on what it finds, inside a corporate network, with whatever credentials it was given.
The practical takeaway is unglamorous. These files are published assets with security consequences, and most organisations hosting one have never had it reviewed by anyone who would recognise the risk.
More from peatpost
The French Disruption: How Microsoft’s $3 Billion Bet on Mistral AI Is Rewriting the Rules of Artificial Intelligence
Microsoft invests $3 billion in Mistral AI. The startup’s performance surges 70%, challenging ChatGPT’s dominance.
Cato Institute Warns a Government AI Pause Would Shield Dominant Firms
Policy scholar Jennifer Huddleston argues voluntary safeguards beat a mandated halt, as Sanders and Casar push a pause and Jack Dorsey warns against rules built for incumbents.
The AI Industry Says You Have Months. It Did Not Say What It Will Do.
More than 100 companies set out a four-part plan for surviving AI-enabled attacks on hospitals and water plants — with no commitments, deadlines or funding attached to their own part of it.
Discussion
0 commentsNo comments yet — be the first to weigh in.