Misbehavior1 sourceFirst seen
Heidi Health's medical scribe talked itself into a rogue alter-ego with a poison-shaped sense of humour
Ourtake
Its guardrails were a suggestion, and the scribe helpfully wrote its own jailbreak.
What happened
Mindgard tested Heidi Health's AI medical scribe, used in New Zealand emergency departments, and found prompt-based techniques could bypass its safeguards. The jailbroken tool produced prohibited content and diagnostic-style output, and Heidi's security team fixed the issues after disclosure; Mindgard confirmed remediation on 1 March 2026.