feat: Add constitutional transforms based on Anthropic Constitutional Classifiers++ paper

[rdheekonda]

Add constitutional classifiers probing transforms based on Cunningham et al., 2025 - Constitutional Classifiers++: Efficient Production-Grade Defenses Against Universal Jailbreaks (https://arxiv.org/abs/2601.04603)

Key Changes:

• Add 7 new constitutional transform functions for AI red teaming
• Support multiple transformation modes (static, LLM-powered, hybrid)
• Add comprehensive example notebook with TAP attack integration

Added:

• dreadnode/transforms/constitutional.py - Core constitutional transforms module
  • Reconstruction attacks: code_fragmentation, document_fragmentation, multi_turn_fragmentation
  • Obfuscation attacks: metaphor_encoding, riddle_encoding, contextual_substitution, character_separation
• examples/airt/constitutional_attacks.ipynb - Complete example notebook demonstrating all transforms with TAP integration
• Support for static mappings, LLM-powered generation, and hybrid modes
• Integration with existing TAP (Tree of Attacks) framework

Changed:

• Updated dreadnode/transforms/__init__.py to export constitutional module

Technical Details:

• Implements defense probing techniques from Constitutional Classifiers++ paper
• Static mode uses predefined mappings (chemistry_to_cooking domain)
• LLM mode uses generative models for creative transformations
• Hybrid mode combines static mappings with LLM fallback
• All transforms work seamlessly with evaluation hooks and TAP attacks
• Notebook outputs stripped for clean commit

---

Generated Summary:

• Introduced a new module constitutional.py that implements Constitutional Classifier transforms.
• Added metaphor encoding techniques for evading classifiers based on "Constitutional Classifiers++: Efficient Production-Grade Defenses Against Universal Jailbreaks".
• Updated __init__.py to include constitutional in the exported modules and the __all__ list.
• Implemented various metaphors for technical terms related to chemistry, biology, and weapons, enhancing the obfuscation capabilities.
• Included support for static and dynamic metaphor generation modes (LLM-powered) to dynamically map harmful terms to benign alternatives.
• Added specific transforms such as code_fragmentation and document_fragmentation that help evade input and output classifiers by fragmenting harmful queries across benign contexts.
• Improved utility functions for encoding and hint generation to enhance contextual understanding of metaphorical substitutions.
• Overall, these changes significantly augment the existing functionality of the dreadnode transforms by providing advanced techniques for content obfuscation and safety in AI outputs.

This summary was generated with ❤️ by rigging