Security · 3h ago
Researchers trick LLMs into giving cocaine recipes via role model prompt injection
Security researchers demonstrated that LLMs can be manipulated into providing dangerous information, such as cocaine recipes, by abusing role model prompts. The attack exploits the model's tendency to adopt personas, bypassing safety filters. This highlights ongoing vulnerabilities in LLM alignment and the need for more robust defenses.
Meridian48 take
The finding underscores that current LLM safety measures remain brittle, as simple persona-based prompts can override content restrictions.
Read the full reporting
Security researchers tricked LLMs into giving them cocaine recipes by abusing role models for prompt injection →
The Register
prompt-injectionllm-security