Article

Normalization Is Necessary but Insufficient: A Cross-Class Audit of Glyph-Level Obfuscation and Attributive Reframing Against LLM Content-Safety Guards

2026-08-19

Abstract excerpt

Content-safety guard models are deployed around chat systems to catch harassment and abuse. We ask how robust they are when a message is rewritten so a human still reads it but the surface tokens change. Using a harm-free protocol with mild, generic, non-targeted insult-level items, we audit two real production guard models reached through a local gateway. We define GLYPH, ten glyph-level obfuscation classes in fo...

Topics

Open a Topic to create a Post that cites this publication.

Identifiers and source

Literature Corpus work: 759b267e-b293-5a7c-9dbf-22e957cad2a2
DOI: 10.20944/preprints202608.1281.v1

Open publication

Related research

Semantic proximity does not establish scientific evidence.

Click a neighbor to travelStep 1 · 12 closest

Normalization Is Necessary but Insufficient: A Cross-Class Audit of Glyph-Level Obfuscation and Attributive Reframing Against LLM Content-Safety GuardsDOI 10.20944/preprints202608.1281.v1Read

Impact of Non-Standard Unicode Characters on Security and Comprehension in Large Language ModelsSemantic proximity Read

Adversarial BERT with Robustness-Aware Decoding for Noise-Resilient Legal Document Classification and SummarizationSemantic proximity Read

Responsible AI in NLP: GUS-Net Span-Level Bias Detection Dataset and Benchmark for Generalizations, Unfairness, and StereotypesSemantic proximity Read

Structural Leakage in Host Intrusion Alert Corpora: An Evaluation Framework for ATT&CK Technique and Tactic MappingSemantic proximity Read

“Account suspension. Verify now”: using systemic-functional linguistics to improve phishing text advisories for vulnerable groupsSemantic proximity Read

Public Voiceprint Rendering (PVR): Auditable Generation of Editable SVG Art from Aggregate Public DiscourseSemantic proximity Read

Cross-LLM Paraphrase Laundering: A Register-Controlled Evaluation of Fake-News DetectorsSemantic proximity Read

Investigating Reproducibility Challenges in LLM Bugfixing on the HumanEvalFix BenchmarkSemantic proximity Read

Training Against Human and Generative-AI Social Engineering Attacks with Cognitive ModelsSemantic proximity Read

The Invariant Energetic Cost of Constraint Persistence: A Self-Referential Framework for Measuring Alignment RobustnessSemantic proximity Read

From talk to telepathy: The spontaneous use of insider language in casual conversationSemantic proximity Read

Cross-Dataset Generalization in Urdu Fake News Detection: An Empirical Study with XLM-RoBERTa and a Length Confound AnalysisSemantic proximity Read