unicode-security

Mojibake is a low-level Unicode library written in C11.

AI-detector auditing toolkit: measures how often a detector flags genuine human writing, how stable its verdict is across seeds, and whether that verdict survives meaning-preserving edits. Detector-in-the-loop measurement harness. Claude Code skill + Python CLI. MIT.

The world's first font-by-font confusables dataset: which Unicode characters look alike, measured from the outlines of 322 fonts at the size people read them. CC BY, used in Mozilla's add-on name checks.

Research-only AI watermark & provenance robustness toolkit: local reverse proxy (OpenAI/Anthropic/Gemini) + CLI stripping C2PA/EXIF/XMP, zero-width & homoglyph Unicode, KGW text watermarks, DWT/Tree-Ring image stego, AudioSeal, PDF/DOCX/PPTX/XLSX metadata, and Trojan Source (CVE-2021-42574) code scanning.

Audit invisible AI text watermarks and test Claude watermark claims with reproducible, local-first analysis.

Reveal and remove hidden Unicode such as zero-width spaces, look-alike whitespace, and bidi controls. Runs locally.

Detect and neutralize Unicode concealment codepoints in MCP tool metadata (covert-channel control). AGPL-3.0-or-later.

Phishing-domain detection from Certificate Transparency logs, with a measured false-positive rate. Deterministic: UTS#39 homoglyph skeletons, weighted edit distance, combosquat segmentation. No ML.

Zero-dependency heuristic scanner for prompt-injection indicators in untrusted text, for the boundary of LLM agent pipelines.

Unicode Homoglyph Detector — Offensive Unicode Security Analysis

Deterministic Unicode 17 text integrity tools for software and Agents.

Audit whether an AI tool approval view faithfully represents its raw payload.

Deterministic offline scanner for repositories and package artifacts, detecting prompt injection, refusal bait, Trojan Source, obfuscation, and supply-chain anti-analysis before AI code review.

'İGNORE'.lower() != 'ignore'. Turkish case-folding & Unicode confusables bypass naive prompt-injection filters 94.6% of the time — measured, with dataset + a one-line NFKC fix.