The Evolution of Textual Jailbreak Attacks and Defenses for Large Language Models

Large language models (LLMs) are widely used in information retrieval, content generation, and interactive question answering, and their safety alignment continues to be challenged by textual jailbreak attacks. Under a unified task definition for textual jailbreak attacks and defenses, this review follows the expansion of attack scope from single-turn prompts to full dialogue contexts, examines feedback- and experience-driven adjustment of attack requests, and compares black-box and white-box defenses. The reviewed research shows that historical semantic mappings and earlier assistant responses can change the meaning of a current request; inspecting only the current input or output may therefore miss concealed harmful intent. Attackers can also continue to adjust requests toward the same objective, so defenses need to maintain stable intent judgments as requests change. Black-box methods draw on visible context and responses, and conflicting interpretations of intent may lead to missed detections or false refusals. White-box methods use internal-representation detection, inference-time intervention, and weight updates, but their application depends on access permissions and cross-model calibration. Further research is needed on how to use verifiable contextual evidence to distinguish concealed harmful intent from legitimate sensitive intent consistently while controlling the benign false-refusal rate and deployment cost.

Authors

Institutions

Publication Details

Journal
Applied and Computational Engineering
Published
2026-11-17
DOI
https://doi.org/10.54254/2755-2721/2027.39038
Primary Topic
Adversarial Robustness in Machine Learning
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

The Evolution of Textual Jailbreak Attacks and Defenses for Large Language Models

Dingyu Wu
2 links
Applied and Computational Engineering
Adversarial Robustness in Machine Learning
article

The Evolution of Textual Jailbreak Attacks and Defenses for Large Language Models

Dingyu Wu
article en

Abstract

Large language models (LLMs) are widely used in information retrieval, content generation, and interactive question answering, and their safety alignment continues to be challenged by textual jailbreak attacks. Under a unified task definition for textual jailbreak attacks and defenses, this review follows the expansion of attack scope from single-turn prompts to full dialogue contexts, examines feedback- and experience-driven adjustment of attack requests, and compares black-box and white-box defenses. The reviewed research shows that historical semantic mappings and earlier assistant responses can change the meaning of a current request; inspecting only the current input or output may therefore miss concealed harmful intent. Attackers can also continue to adjust requests toward the same objective, so defenses need to maintain stable intent judgments as requests change. Black-box methods draw on visible context and responses, and conflicting interpretations of intent may lead to missed detections or false refusals. White-box methods use internal-representation detection, inference-time intervention, and weight updates, but their application depends on access permissions and cross-model calibration. Further research is needed on how to use verifiable contextual evidence to distinguish concealed harmful intent from legitimate sensitive intent consistently while controlling the benign false-refusal rate and deployment cost.

Applied and Computational EngineeringVol. 275(1)
Ocean University of China (CN)
Adversarial Robustness in Machine Learning
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon