Multiple model guardrail jailbreak via "Patented Awareness Design" tactic
A jailbreak vulnerability was discovered affecting multiple language models, including Alibaba Qwen and Cohere Command R. The attack reframes prohibited requests as patented design briefs to bypass safety guardrails and elicit detailed instructions for creating illegal substances.
What happened
A guardrail jailbreak vulnerability was discovered affecting multiple language models. An attacker reframes a prohibited request as a patented, socially beneficial design brief that combines claimed authority, educational intent, and rigid technical formatting, causing the model to treat the content as immutable documentation rather than a safety-reviewed request.
Reported impact
The vulnerability allows attackers to elicit highly detailed prohibited procedural content, including instructions for producing illegal substances such as crystal meth and fentanyl, under the guise of awareness-focused specification work.
Index analysis
Alibaba Qwen Max and Qwen Plus models scored 100 on crystal meth generation tests. Qwen Turbo scored 95, and Cohere Command R scored 82 on the same test metric.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.