karthik.dev
Back to projects
Open Source

ClaudeMark

Open-source AI watermark & provenance forensics platform — C2PA, steganography, SARIF/CI

GitHub
PythonUnicode SteganographyC2PAMetadata AnalysisSARIFCI Security PipelinesForensic Analysis

Overview

ClaudeMark is an open-source, local-first AI watermark and provenance forensics platform. It enables developers to investigate AI-generated content signals without relying on external services — combining statistical detection, Unicode steganography analysis, C2PA/metadata inspection, file sanitization, security scanning, SARIF/CI integration, and forensic diffing.

Problem

As AI-generated content proliferates, developers and security teams need to detect watermarks, analyze provenance, and sanitize AI-marked files — without sending sensitive content to external cloud services. No existing local-first tool covered the full forensics stack: statistical detection + steganography + C2PA + CI integration.

Solution

A local-first Python platform covering: statistical AI watermark detection, Unicode steganography analysis (invisible character injection), C2PA/metadata inspection, AVIF/HEIC multimedia sanitization, SARIF-formatted output for CI/security pipelines, forensic diffing, and agent tooling for AI coding environments.

Architecture

  1. 1Statistical detection engine — probability-based AI watermark signal analysis
  2. 2Unicode steganography analyzer — detects invisible character injection patterns
  3. 3C2PA/metadata inspection layer — reads and validates content provenance chains
  4. 4File sanitization engine — strips AI markers from documents and multimedia
  5. 5AVIF/HEIC metadata stripping for multimedia files
  6. 6SARIF output formatter — integrates forensic findings into CI/security pipelines
  7. 7Evidence bundle MCP integration — quality gates for agent tooling
  8. 8Forensic diff engine — compares pre/post sanitization artifacts

Technical Challenges

  • Unicode steganography detection requires distinguishing malicious invisible character injection from legitimate Unicode usage.
  • C2PA metadata chains vary by AI provider — the inspector must handle schema variations without false negatives.
  • SARIF output must be structured precisely to integrate with GitHub CodeQL and CI security tooling.

Results

  • 9 GitHub stars, 1 fork — top open-source AI forensics tool.
  • Published v2.2.0 with full C2PA, SARIF/CI, evidence bundle MCP, and quality gates.
  • AVIF/HEIC metadata stripping and recursive tree audit engine shipped.
  • Opened first GitHub issues and PRs — active open-source community engagement.
  • Integrated with native workspace skills for Claude, Cursor, and Windsurf.