Linguistic Pattern Obfuscation For Detection Evasion

via “invisible unicode and encoding-based obfuscation detection”

Open-source LLM input/output security scanner toolkit.

Unique: Specialized detection for unicode-based obfuscation techniques (zero-width characters, homoglyphs, combining marks) that other scanners may miss; analyzes character encodings at the unicode level rather than semantic level; prevents evasion of other security scanners through encoding tricks

vs others: More targeted than generic text sanitization because it specifically detects obfuscation patterns; complements other scanners by catching evasion attempts that use unicode tricks; runs locally with no external dependencies

via “obfuscation detection and deobfuscation assistance”

Show HN: Ghidra MCP Server – 110 tools for AI-assisted reverse engineering

Unique: Combines pattern detection with heuristic analysis to identify obfuscation techniques and provide deobfuscation guidance, rather than just flagging suspicious code

vs others: Provides actionable deobfuscation hints alongside detection, enabling LLMs to assist in understanding obfuscated code

via “prompt-obfuscation-and-evasion-technique-catalog”

LEAKED SYSTEM PROMPTS FOR CHATGPT, CLAUDE, GEMINI, GROK, PERPLEXITY, CURSOR, LOVABLE, REPLIT, AND MORE! - AI SYSTEMS TRANSPARENCY FOR ALL! 👐

Unique: Documents obfuscation techniques (leetspeak, special characters, context manipulation) as reproducible attack patterns with model-specific effectiveness data, rather than treating them as one-off exploits. The repository tracks which obfuscation strategies work against which models and versions.

vs others: Provides a curated, model-specific catalog of obfuscation techniques with effectiveness metrics, whereas most security research on prompt injection evasion is scattered across informal disclosures without systematic evaluation.

via “multi-language-content-rewriting-with-cross-lingual-evasion”

Unique: Applies language-specific obfuscation patterns that account for grammatical structures and morphological variations unique to each language, rather than using language-agnostic paraphrasing; likely maintains separate detection signature models per language to account for language-specific detection heuristics

vs others: Handles non-English content with language-aware transformations vs. generic paraphrasing tools that treat all languages identically, but support is limited to major languages and detection evasion effectiveness varies significantly by language

via “detection evasion through linguistic transformation”