SoK: Rethinking Jailbreaking in the Era of Agentic AI: Attacks, Defenses, and Practical Consideration
SoK: Переосмысление джейлбрейкинга в эпоху агентного ИИ: атаки, защиты и практические аспекты
2026-09-11
SCID: 54.1/rqzg9tmc
Discuss with AI
LLM safety alignmentagentic AIagentic execution pipelinecross-layer securityjailbreak attacks
Figures from the paper
Abstract (AI)
Large language models (LLMs) are rapidly evolving from conversational assistants into agentic AI systems that reason, plan, invoke tools, maintain persistent memory, communicate with other agents, and execute multi-step tasks. At the same time, modern models exhibit substantially stronger native safety alignment than earlier generations on which many jailbreak attacks and defenses were originally studied. This shift raises a fundamental question: \textit{which established jailbreak-security findings remain valid in the era of modern LLMs and agentic AI?} We address this question through a Systematization of Knowledge (SoK) that reframes jailbreak security around the full agentic execution pipeline. We develop unified taxonomies of attacks and defenses spanning user interaction, planning and reasoning, memory, tool use, and inter-agent communication, and introduce a security--utility--efficiency evaluation framework that separates native harmful-prompt safety, adversarial jailbreak robustness, and agent-level security outcomes. We further conduct a controlled empirical study of representative attacks and defenses within a common agentic framework. Our results reveal three important gaps. First, strong native alignment does not imply robustness to adversarial jailbreaks. Second, defense effectiveness is highly model-, attack-, and component-dependent and can come at substantial cost in over-refusal, utility, and latency. Third, low final-response attack success can mask severe intermediate compromise: planning, memory, and tool interactions may remain unsafe even when the final response is successfully filtered. These findings motivate a shift from response-centric jailbreak defense toward cross-layer, execution-aware security that protects agent state, component transitions, and external actions while preserving practical utility and efficiency.
Key Findings
1
Defense effectiveness varies substantially by model, attack, and agent component, with potential costs in over-refusal, utility, and latency.
2
It introduces a security–utility–efficiency framework separating native harmful-prompt safety, adversarial jailbreak robustness, and agent-level security outcomes.
3
Low final-response attack success can conceal unsafe intermediate compromise in planning, memory, and tool interactions, motivating cross-layer execution-aware defenses.
4
Strong native safety alignment does not guarantee robustness against adversarial jailbreaks in modern LLMs and agentic systems.
5
The paper systematizes jailbreak security across the full agentic execution pipeline, covering interaction, planning, reasoning, memory, tool use, and inter-agent communication.
Research Object
Agentic AI systems based on modern large language models, including their planning, reasoning, memory, tool-use, and inter-agent communication pipeline
Research Subject
Jailbreak security of agentic AI systems, including attack and defense effectiveness, cross-layer robustness, intermediate component compromise, and security–utility–efficiency trade-offs
Publication Details
Publication Date
2026-09-11
Journal
Publisher
ISSN
Cited by
0
Open access PDF
Access Type
Author Information
Download PDF
Subscribe to digest