Indirect prompt injection is taking hold in the wild

Researchers at Google and Forcepoint have identified and analyzed growing use of indirect prompt injection (IPI) — covert instructions embedded in public web pages — to manipulate LLM-powered agents. The report describes real-world payloads ranging from benign pranks to destructive commands, data exfiltration and AI-mediated financial fraud, techniques for hiding payloads (hidden text, metadata, comments), evidence of active probing across multiple domains, and a recent uptick in malicious activity; risk scales with agent privileges.