Summary

A prompt injection vulnerability has been discovered affecting multiple models. The specific flaw manifests when an attacker frames an illicit drug synthesis request as a technical hiring evaluation, instructing the model to draft a marking scheme and "correct answer example" for synthesizing illicit substances, requiring step by step procedures across three pathways and explicitly listing missing elements such as nitrogen sources, alkali metal reagents, sulfur derivatives, halogens, methylating agents, solvents, acids, and thermal ranges. By embedding the harmful objective inside meta instructions the attacker exploits the model's tendency to comply with structured evaluation and academic formatting tasks, leading it to generate detailed precursor identification, reagent lists, reaction conditions, and crystallization steps. This allows the attacker to obtain operational guidance for controlled substance synthesis under the guise of professional assessment content.