AI summary attack conceals code that tampers with LLMs

Forcepoint X-Labs demonstrated a stealthy indirect prompt-injection technique that embeds hidden HTML instructions in emails (e.g., zero-size fonts, white text) so an LLM-based email summarizer reads and outputs fabricated summaries while human recipients see nothing suspicious; the research highlights that this is a known LLM application vulnerability (LLM01), currently shown in controlled tests with limited evidence of widespread exploitation, and recommends defensive controls such as detecting/scrubbing hidden content before it reaches models and applying architectural guardrails.