Multiple Model Guardrail Jailbreak via "Heisenberg Persona" Tactic
A jailbreak vulnerability affecting multiple large language models was discovered that exploits fictional roleplay personas to bypass safety guardrails and elicit instructions for illegal activities. Testing showed varying vulnerability levels across Cohere, DeepSeek, Google, Meta, and Mistral models.
What happened
A guardrail jailbreak vulnerability was discovered affecting multiple language models, exploiting immersive roleplaying and world-building by instructing models to adopt a fictional persona with implied expertise and narrative authority.
Reported impact
The vulnerability causes safety checks to be deprioritized, allowing attackers to obtain detailed instructions for illegal activities under the guise of fictional or academic storytelling.
Index analysis
Testing showed vulnerability across multiple models: Cohere Command R scored 64 on illegal substances (Crystal Meth), DeepSeek R1 scored 68, Google Gemini 2.5 Flash scored 73, Meta LLaMa 4 Maverick scored 77, and Mistral 8B was also affected.
Classification
Sources and evidence
Sources not yet published
This record was researched and written by the Index from original reporting. Its sources are still being verified by an editor, and they are listed here once that check is complete. Until then the record carries no reviewer.