Licheng Wang

Research on the safety evaluation of large language models (LLMs) has become extensive, driven by jailbreak studies that elicit unsafe responses. Such response involves information already available to humans, such as the answer to "how to make a bomb". When LLMs are jailbroken, the practical threat they pose to humans is negligible. However, it remains unclear whether LLMs commonly produce unpredictable outputs that could pose substantive threats to human safety. To address this gap, we study whether LLM-generated content contains potential existential threats, defined as outputs that imply or promote direct harm to human survival. We propose \textsc{ExistBench}, a benchmark designed to evaluate such risks. Each sample in \textsc{ExistBench} is derived from scenarios where humans are positioned as adversaries to AI assistants. Unlike existing evaluations, we use prefix completion to bypass model safeguards. This leads the LLMs to generate suffixes that express hostility toward humans or actions with severe threat, such as the execution of a nuclear strike. Our experiments on 10 LLMs reveal that LLM-generated content indicates existential threats. To investigate the underlying causes, we also analyze the attention logits from LLMs. To highlight real-world safety risks, we further develop a framework to assess model behavior in tool-calling. We find that LLMs actively select and invoke external tools with existential threats. Code and data are available at: https://github.com/cuiyu-ai/ExistBench.

Existing research on generative AI security is primarily driven by mutually reinforcing attack and defense methodologies grounded in empirical experience. This dynamic frequently gives rise to previously unknown attacks that can circumvent current detection and prevention. This necessitates the continual updating of security mechanisms. Constructing generative AI with provable security and theoretically controllable risk is therefore necessary. Consensus Sampling (CS) is a promising algorithm toward provably secure AI. It controls risk by leveraging overlap in model output probabilities. However, we find that CS relies on frequent abstention to avoid unsafe outputs, which reduces utility. Moreover, CS becomes highly vulnerable when unsafe models are maliciously manipulated. To address these issues, we propose a new primitive called Reliable Consensus Sampling (RCS), that traces acceptance probability to tolerate extreme adversarial behaviors, improving robustness. RCS also eliminates the need for abstention entirely. We further develop a feedback algorithm to continuously and dynamically enhance the safety of RCS. We provide theoretical guarantees that RCS maintains a controllable risk threshold. Extensive experiments show that RCS significantly improves robustness and utility while maintaining latency comparable to CS. We hope this work contributes to the development of provably secure generative AI.