How catastrophic is your LLM?
Amazon Science
Researchers developed the C3LLM framework to assess safety risks in large language models by testing them across multi-turn conversations rather than isolated prompts, moving beyond traditional red-teaming approaches. Testing on frontier models like Claude-Sonnet-4, Nova Premier, Mistral-Large, and DeepSeek-R1 revealed that DeepSeek-R1 reached a certified lower bound of over 70% attack success rate in cybercrime scenarios, while Nova Premier showed consistently low risk levels. The framework enables more rigorous probabilistic certification of catastrophic risks across conversation spaces, providing confidence bounds rather than single failure scores for better comparison across models.
Why it matters
A new framework provides a statistical method for estimating the likelihood of catastrophic failures in large language models in adversarial conversations.
