Tag
A practical walkthrough on how to test AI agents with multi-turn attacks to uncover vulnerabilities that seem harmless in isolation but become critical over interactions.
The paper introduces Blueprint, a safety-evaluation framework that uses WorldviewSim and Monte Carlo Tree Search to optimize multi-turn jailbreak attacks against large language models, achieving high attack success rates with few queries and revealing model-specific vulnerabilities.
EvoFlint introduces an evolutionary quality-diversity search method for multi-turn red-teaming of large language models, aiming to map vulnerabilities rather than just break models, achieving high attack success rates on various LLMs.
This paper introduces Temporal Context Awareness (TCA), a defense framework that detects multi-turn manipulation attacks on LLMs by analyzing semantic drift, cross-turn intention consistency, and evolving conversational patterns to mitigate adversarial context-building across dialogues.
This paper proposes Dialogue Critic Guided Sampling (DCGS), a framework that defends LLMs against multi-turn adversarial attacks by inferring user intent from conversation history and using value/regret-based critics to score responses, achieving improved robustness without fine-tuning.
A six-month analysis of real adversarial inputs reveals that simple multi-turn setups, forward-momentum exploitation, and role redefinition attacks consistently bypass single-message classifiers. The post argues that stateful monitoring of conversational context is more effective than improving one-shot detection.
An attacker can bypass security by spreading malicious instructions across multiple messages; Bendex Arc is a tool that tracks session behavior across turns to catch such attacks.