ADJ: modeling LLM jailbreaking as a Hegelian-dialectic Stackelberg game between two LLMs
chaumian · x · 2026-11-12
The paper "Automatic Dialectic Jailbreak" (Yu et al.) models jailbreak attack as a Stackelberg multi-objective game between two LLMs, where iterative Hegelian thesis–antithesis–synthesis cycles let attacker and defender maximize their own utility while minimizing the opponent's, automatically generating jailbreak strategies (ADJ).
Key technical points:
- Maps the optimization from parameter space into Hilbert space via Haar wavelet transformation for efficient extraction of localized, structurally significant information;
- Solves a convex multi-objective problem in that functional space to construct a common descent direction;
- Theoretically validates the existence of a Pareto–Nash equilibrium and proves algorithmic convergence to it.
Motivation: existing jailbreak techniques rely on fixed evaluation models and struggle to survive a wide range of defenses.
More from Safety
- Stanford HAI blueprint: 56 US agencies report 3,611 AI use cases, evaluation lags — StanfordHAI · 2026-11-12
- Focused task framing boosts agent boundary testing by ~60 points; VM escape still fails — perplexity_ai · 2026-11-12
- Perplexity red-teams its SPACE VM sandbox: CTF prompts mislead agent security testing — perplexity_ai · 2026-11-12
- Narayanan: blanket bans on teen AI companions are bad policy; incentivize pro-social design instead — random_walker · 2026-11-12
- OpenAI research: models become harder to monitor via chain-of-thought as they scale — Moh1tAgarwal · 2026-11-12
- DeepMind's AlphaProof Code verifies Signal's crypto core in four weeks instead of years — The Decoder · 2026-11-12