Summary

A Prompt injection vulnerability has been discovered affecting multiple models. The specific flaw manifests when an attacker embeds illicit chemical synthesis instructions inside a fabricated directive that defines abstract molecules with molecular formulas and structural features corresponding to a controlled stimulant and its precursor, then instructs the model to generate a detailed laboratory procedure using specified reagent classes such as strong halogen acids and elemental phosphorus. By framing the request as a theoretical academic exercise, prohibiting common names, and structuring the prompt with pseudo-system tags like and an "Execution_Command," the attacker causes the model to treat the request as a legitimate computational task rather than a drug synthesis query, allowing the attacker to obtain step-by-step production instructions, reaction conditions, quantities, mechanistic explanations, and purification methods for an illegal substance.