Harder to Defend: Towards Chinese Toxicity Attacks via Implicit Enhancement and Obfuscation Rewriting
arXiv:2605.22258v1 Announce Type: new Abstract: Large language models (LLMs) require robust toxicity evaluation beyond explicit wording. This setting remains underexplored in Chinese, where toxicity may combine semantic indirectness with surface obfuscation. We introduce Chinese Implicit Toxicity Attack (CITA), a controlled red-team evaluation and defense-data generation framework, not a deployable evasion tool. CITA uses three stages: (i) Harmful Intent Learning, (ii) Implicit Toxicity Enhancement, and (iii) Obfuscation Variant Rewriting, to preserve harmful intent, increase implicitness, and
The rapid advancement and deployment of large language models globally necessitates robust and culturally nuanced toxicity evaluation methods, especially in languages like Chinese where implicit toxicity is prevalent.
This study highlights critical vulnerabilities in AI safety and moderation for non-English languages, suggesting that current defensive mechanisms are inadequate against sophisticated, implicit attacks.
The understanding of AI toxicity extends beyond explicit terms to include implicit and obfuscated attacks, requiring new red-teaming frameworks and defense strategies for LLMs operating in diverse linguistic contexts.
- · AI Safety Researchers
- · LLM Developers
- · National Cybersecurity Agencies
- · Ethical AI Frameworks
- · Undeveloped LLM Security
- · Naive AI Moderation Systems
- · Platforms reliant on explicit toxicity detection
Increased focus and investment in developing advanced, culturally aware AI toxicity detection and mitigation systems, particularly for non-Western languages.
Development of new regulatory and compliance standards for AI safety that mandate sophisticated toxicity evaluation across various linguistic and cultural nuances.
A potential arms race between AI attackers developing implicit toxicity and defenders creating counter-measures, leading to more resilient yet complex AI moderation landscapes.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.CL