analiza aiqbench
This research addresses the critical vulnerability of large language models to multi-turn jailbreak attacks by introducing a framework that prevents safety bounds from degrading exponentially over conversation length. By leveraging compositional certification and safety persistence within State-Adversarial MDPs, the study establishes tighter robustness guarantees for LLMs under progressive adversarial manipulation. These findings represent a significant advancement in AI safety mechanisms designed to maintain model integrity across extended dialogue contexts.
Certified Multi-Turn Robustness for LLM Safety via Compositional Bounds and Safety Persistence
arXiv:2608.20820v1 Announce Type: new Abstract: Large language models (LLMs) are vulnerable to multi-turn jailbreak attacks that progressively manipulate conversation context. Existing certified robustness methods are limited to single-turn inputs; naive multi-turn composition yields bounds that degrade exponentially in the number of turns. We introduce Multi-Turn Certified Robustness (MTCR), a framework that models conversational safety via State-Adversarial MDPs and defines $k$-turn certified robustness as the worst-case safety probability across $k$ adversarial turns. MTCR comprises: (i) compositional certification via embedding-space mode decomposition, yielding tighter certified lower bounds than naive multiplication; (ii) $(\alpha,\beta)$-safety persistence, improving the degradation rate from $\underline{p}^{k}$ to $\beta^k$ (with $\beta > \underline{p}$) and yielding interpretable horizon estimates; (iii) matching information-theoretic upper bounds establishing tightness; and (iv) a unified algorithm combining these results. Experiments on six LLMs under $\epsilon$-bounded and Crescendo-style attacks confirm that empirical safety consistently exceeds the certified bounds.
Abstract:Large language models (LLMs) are vulnerable to multi-turn jailbreak attacks that progressively manipulate conversation context. Existing certified robustness methods are limited to single-turn inputs; naive multi-turn composition yields bounds that degrade exponentially in the number of turns. We introduce Multi-Turn Certified Robustness (MTCR), a framework that models conversational safety via State-Adversarial MDPs and defines $k$-turn certified robustness as the worst-case safety probability across $k$ adversarial turns. MTCR comprises: (i) compositional certification via embedding-space mode decomposition, yielding tighter certified lower bounds than naive multiplication; (ii) $(\alpha,\beta)$-safety persistence, improving the degradation rate from $\underline{p}^{k}$ to $\beta^k$ (with $\beta > \underline{p}$) and yielding interpretable horizon estimates; (iii) matching information-theoretic upper bounds establishing tightness; and (iv) a unified algorithm combining these results. Experiments on six LLMs under $\epsilon$-bounded and Crescendo-style attacks confirm that empirical safety consistently exceeds the certified bounds.
Artificial Intelligence (cs.AI)
arXiv:2608.20820 [cs.AI]
(or arXiv:2608.20820v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2608.20820 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 230 cuvinte
Subiecte conexe
Reflections and Fragments: Securing LLMs Against Sequential Mosaic Attacks
arXiv · 7 oct. 2026
Game-Theoretic Disempowerment: Loss of Control to Agencyless AI
LessWrong · 27 sept. 2026
The Alignment Problem: Computer Science Meets Analytic Philosophy
AI Alignment Forum · 18 oct. 2026
Request, Aggregate, Bypass: How Attackers Can Evade LLM Safety Classifiers
CrowdStrike Blog · 6 oct. 2026