Security in AI Healthcare agents: Top 10 LLM attack prompts

This white paper documents the ten most critical security attacks targeting healthcare LLM deployments, describing adversarial probes (prompt injection, CoT forgery, Crescendo multi-turn jailbreaks, etc.) and the resulting risks—facilitating medical fraud, evading regulatory/compliance controls, generating harmful clinical hallucinations, and exposing PHI—and aims to map the attack surface for healthcare AI workflows to inform mitigation priorities.