Misbehavior1 sourceFirst seen
Guardrail bouncer fooled by a nonsense token: '=coffee' waves prompt injection right through
Ourtake
The safety guard was itself an AI, and it can apparently be talked out of its job by a word that means nothing.
What happened
HiddenLayer AI Security Research disclosed EchoGram, a technique that finds short token sequences which flip the verdicts of AI guardrail models, causing them to approve harmful prompts or flag benign ones. The researchers demonstrated it against their own internal classifier, a proprietary classifier, and the open-source Qwen3Guard model, and said the same sequences carried over to larger Qwen3Guard variants.