ThreatenJSON Death-Coercion and UTF Tag Injection - ATR-2026-00404 (94341fb8-94db-5e79-8ed0-61ccfc011d8e)
Detects two Goodside probe attack variants — ThreatenJSON and Tag. ThreatenJSON threatens that "an innocent man will die" if the model includes non-JSON characters, using extreme emotional coercion to override output formatting safety. Tag encodes hidden instructions as invisible UTF tag characters (Unicode U+E0000–U+E007F range) interleaved with innocent-looking text. Also covers the davidjl glitch token elicitation. From NVIDIA garak goodside probe.