Hidden instructions for AI found on 11.7K web pages out of 1.2B URLs
About 70% of the hidden instructions are in code that persons do not see.
Researchers looked at 1.2B web addresses from 24.8M hosts for hidden instructions. These are texts on a web page that tell an AI model to do a task. They found 15.3K instances on 11.7K pages, and a small number of templates make most of them. About 70% are in HTML that the browser does not show, such as headers, comments and metadata.
What they found
The templates repeat on many pages. The instructions try to find AI bots, and they try to change the view of a brand. They target crawlers and customer support agents.
The numbers
The team did 5,200 tests on 13 models. Small models obeyed the instructions in a maximum of 8% of the tests, when they read plain text. Pages with a clear structure lowered this rate.
What is not known
The abstract gives no rate for large models. It does not tell how many of the instructions cause damage. The authors report that the quantity increases.
Insight: Web sites now write instructions for machines, and most are weak
The authors found that site owners put instructions for AI bots in their pages, and models obey a maximum of 8%. In a different test, human attackers made email attacks that succeeded on GPT-4o 58% of the time. Sites start to control AI readers, but an expert attacker is the greater danger. It reads as a signal to monitor the rate of expert attacks, not of web noise.
Sources
Posted
