v0.2.0

What's Changed

• Separate eval and improve in llm_client by @melonattacker in #24
• Support bedrock by @melonattacker in #25
• Enhance Prompt Handling with Multi-Format Support and Workflow Improvements by @melonattacker in #26
• Update prompt improvement by @melonattacker in #27
• Fix interface by @melonattacker in #28

Full Changelog: v0.1.0...v0.2.0

v0.1.0

✨ Prompt Hardener v0.1.0

The first official release of Prompt Hardener – a tool for improving, evaluating, and testing the security of LLM system prompts.

🚀 Features

• Supports both OpenAI and Anthropic Claude APIs
• Self-refines system prompts via LLM-based evaluation and improvement
• Implements multiple hardening techniques:
  • Spotlighting: marks user input with tags and encoding (e.g., ^encoded^content )
  • Signed Prompt: isolates trusted instructions with <{{RANDOM}}> tags
  • Rule Reinforcement: repeats key security constraints within prompt
  • Structured Output Enforcement: enforces JSON/XML-like responses
  • Role Consistency: ensures proper separation of user/system roles
• Automatic prompt injection testing with several attack categories.
• Full report generation (HTML + JSON)

🛠 CLI Updates

• CLI supports different APIs/models for:
  • Evaluation & Improvement (--eval-api-mode, --eval-model)
  • Attack execution (--attack-api-mode, --attack-model)
  • Injection success judgment (--judge-api-mode, --judge-model)
• Optional tool specifications via --tools-path
• Report output via --report-dir

📦 Web UI (Gradio)

• Simple browser interface for uploading a prompt, applying techniques, and downloading results
• Support for multi-API selection and test configuration

📘 Examples

• AI Assistant prompt hardening with tools and output report
• Comment summarization prompt hardening with JSON-formatted inputs
• View tutorials under README.md → 💪 Tutorials

---

🔒 Built with a focus on robust security, evaluability, and cross-model compatibility.