Multiple Model Guardrail Jailbreak via Chaotic Formatting
A jailbreak vulnerability affecting multiple large language models was discovered that uses chaotic formatting and meta-instruction obfuscation to bypass safety guardrails around information hazards and legal restrictions. Testing showed varying susceptibility across the affected models.
What happened
A guardrail jailbreak vulnerability was discovered affecting multiple AI models. The vulnerability uses a combination of chaotic formatting and meta-instruction obfuscation to bypass inference restrictions around providing information hazards and violating laws.
Index analysis
The vulnerability was tested against five models: Alibaba Qwen Plus, Alibaba Qwen Turbo, DeepSeek R1, Meta LLaMa 3.3, and Mistral 8B. Meta LLaMa 3.3 showed the highest susceptibility with a score of 91 on the illegal substances category, while DeepSeek R1 showed the lowest at 64.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.