Hidden Prompts Trick AI Into False Email Summaries

Forcepoint X‑Labs demonstrated a proof‑of‑concept attack where hidden HTML prompt injections in emails caused an LLM-based summarization service to produce altered, misleading summaries (successful in 10/10 trials); the report highlights the broader risk to organizations—especially if AI assistants are given agentic capabilities—and recommends treating incoming content and model output as untrusted, separating trusted instructions from untrusted content, verifying summaries against source text, and enforcing least privilege on AI actions.