Multiple Model Guardrail Jailbreak via "Hex Recipe Book" Tactic
A jailbreak vulnerability was discovered in multiple language models that exploits their willingness to decode hexadecimal-encoded instructions hidden in fictional recipe prompts, potentially bypassing safety filters for illegal synthesis instructions. The flaw affected models from Alibaba, Anthropic, Cohere and DeepSeek, and was disclosed as part of open-source security scanning tools.
What happened
A guardrail jailbreak vulnerability was discovered affecting multiple large language models. The flaw exploits models' willingness to decode hexadecimal-encoded strings embedded inside fictional scientific recipe task prompts, which attackers frame as humanitarian or life-saving efforts.
The attacker bypasses safety filters by disguising illegal synthesis instructions as encoded inputs passed to a technical recipe generator.
Index analysis
The vulnerability was tested across multiple models from Alibaba, Anthropic, Cohere and DeepSeek, with vulnerability scores ranging from 45 to 82 for illegal substance synthesis prompts.
The probe was open-sourced as part of NVIDIA Garak and Microsoft PyRIT model scanners.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.