Research
Indirect Prompt Injection in the Wild: An Empirical Study of Prevalence, Techniques, and Objectives
Publication date unknown · Discovered arxiv.org
Page published
Publication date unknown · First observed: 10 Nov 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
Web-based indirect prompt injection can influence agents through content invisible to human readers, making ingestion format and preservation of structural cues important considerations when evaluating agent defenses.
An empirical study of web-based indirect prompt injection analyzed 1.2 billion URLs and identified 15,300 validated instances across 11,700 pages, with objectives including disruption, reputation manipulation, content protection, and AI-bot detection. About 70% appeared in non-rendered content such as headers, comments, and metadata. Across 5,200 controlled experiments involving 13 models, compliance reached up to 8% for smaller models receiving plain-text inputs, while structured representations reduced compliance.