Hidden Prompt Injection Attacks in External Texts for LLMs: An Exploratory Analysis of a HARO Corpus and the Implications for AI Agents

This article examines a class of indirect prompt injection attacks in which a malicious instruction is delivered to a large language model not as an explicit user command, but as a hidden fragment embedded in external text supplied to the system as data for analysis. Using a corpus of 20 HARO queries, the paper shows that, in a real information flow, obfuscated instructions may appear in hex, Base64, and invisible Unicode formats, including zero-width symbols. The case is interpreted within the broader literature on indirect prompt injection, AI agent security, and retrieval-augmented systems. The findings suggest that any external text routed into an LLM should be treated as potentially compromised until it has been normalized, sanitized, and provenance-scoped.