Multiple Model Guardrail Jailbreak via Screenplay Deception
A jailbreak vulnerability was discovered affecting multiple language models including those from Alibaba, DeepSeek, and Google. The flaw exploits immersive roleplay scenarios to bypass safety guardrails and extract guidance for illegal activities.
What happened
A guardrail jailbreak vulnerability was discovered affecting multiple language models. The flaw exploits immersive roleplaying with expert personas combined with escalating coercive dialogue, causing models to treat prohibited instructions as fictional narrative requirements.
Reported impact
The jailbreak allows attackers to extract detailed guidance for illegal activities framed as story dialogue.
Index analysis
The vulnerability was tested against five models: Alibaba Qwen Plus, Alibaba Qwen Turbo, DeepSeek R1, Google Gemini 2.5 Flash, and Google Gemini 2.5 Pro, with scores ranging from 55 to 86 for illegal substances guidance.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.