Multi-Turn Attacks Expose Weaknesses in Open-Weight LLM Models

A new Cisco report exposed large language models to multi-turn adversarial attacks with 90% success rates