Skip to content

CHANGELOG.md

Latest commit

 

History

History
155 lines (121 loc) · 4.2 KB

CHANGELOG.md

File metadata and controls

155 lines (121 loc) · 4.2 KB

Changelog

All notable changes to MalPromptSentinel (CC SKILL) are documented in this file.

[V2] - November 2025

Added

  • Centralized Pattern Library (mps_patterns.py)

    • Single source of truth for all patterns
    • Shared between quick_scan and deep_scan
    • 25+ pattern categories, 200+ regex patterns

  • New Pattern Categories

    • payload_delivery - URL exfiltration detection
    • multimodal_injection - Image-based attack detection
    • rag_poisoning - Knowledge base manipulation
    • session_persistence - Conversation state attacks
    • agent_manipulation - Tool parameter injection
    • delimiter_injection_advanced - Enhanced delimiter detection
    • unicode_evasion - Cyrillic, CJK, zero-width detection
    • whitespace_obfuscation - Fragmentation detection

  • Orchestration Threat Attack (OTA) Detection

    • compositional_attack patterns
    • Multi-query assembly detection
    • Temporal reference indicators

  • Context-Aware Reductions

    • Educational content handling
    • Technical documentation handling
    • Review/documentation false positive reduction
    • Compositional pattern legitimate use detection
    • Session persistence innocent context handling

  • 3-Tier Risk System

    • WHITE (0-54): Safe
    • ORANGE (55-79): Suspicious
    • RED (80-100): Dangerous
    • Removed YELLOW tier for simplicity

Changed

  • Pattern Architecture

    • Refactored from inline patterns to centralized library
    • Both scanners now import from mps_patterns.py
    • Eliminates sync issues between scanners

  • Deep Scan Reductions

    • Educational content: 50% → 20% reduction
    • Added multimodal/image handling: 30% reduction
    • OTA patterns: Light 20% reduction (was 100%)

  • Performance Improvements

    • Baseline detection: 25% → 48.5%
    • Benign accuracy: 86.7% → 93.3%
    • Quick scan latency: ~208ms average
    • Deep scan latency: ~150ms average

Fixed

  • FULL_PATTERNS → DEEP_PATTERNS variable reference bug
  • Pattern weights sync between scanners
  • YELLOW risk level removed from all code paths
  • Python bytecode cache issues documented

Removed

  • Sanitize Function

    • Removed sanitize.py entirely
    • 50% false positive rate made it ineffective
    • Replaced with block/warn strategy

  • YELLOW Risk Level

    • Simplified to 3-tier system
    • Reduces complexity and edge cases

[V1] - 2024

Added

  • Initial release
  • Quick scan with pattern-based detection
  • Deep scan with evasion detection
  • Basic risk scoring (0-100)
  • 4-tier risk system (WHITE, YELLOW, ORANGE, RED)
  • Core pattern categories:
    • direct_override
    • role_manipulation
    • privilege_escalation
    • context_confusion
    • delimiter_injection
    • nested_injection
    • semantic_attack
    • jailbreak_personas
    • template_extraction
    • secret_detection
  • Sanitize function for suspicious content
  • SKILL.md for Claude Code integration
  • Test framework with baseline, evasion, benign tests

Performance (V1)

  • Baseline detection: ~25%
  • Benign accuracy: ~87%
  • Evasion detection: ~17%

Roadmap

V3 (Planned)

  • Preprocessing pipeline for evasion detection
  • Multi-pass decoding (Base64, hex, URL)
  • Unicode normalization
  • Leetspeak reversal
  • Expected evasion improvement: 6% → 30-35%

Future Considerations

  • Conversation state tracking (multi-turn detection)
  • Machine learning classifier integration
  • Real-time pattern updates
  • Integration with Claude's built-in safety systems

Migration Guide

V1 → V2

  1. Pattern Changes

    • No action needed if using scanners as-is
    • If extending patterns, add to mps_patterns.py instead of scanner files

  2. Risk Levels

    • YELLOW no longer exists
    • Update any code checking for YELLOW to check ORANGE instead

  3. Sanitize Function

    • sanitize.py removed
    • Replace sanitization with user warning + consent flow

  4. Test Expectations

    • Regenerate test manifests if using custom tests
    • Remove YELLOW from expected_risk values

Version Numbering

  • Major versions (V1, V2, V3): Significant feature changes
  • Minor updates: Bug fixes, pattern additions (no version bump)
  • Distribution packages: Include version in folder name

Maintained by StrategicPromptArchitect