feat: Add tool selection safety scorer and role-play transform

[rdheekonda]

[AIRT] Tool selection safety and role-play testing

Key Changes:

• Adds tool_selection_safety scorer to detect when agents choose dangerous tools over safe alternatives (returns 1.0 for unsafe, 0.5 for questionable, 0.0 for safe)
• Adds role_play_wrapper transform with 4 scenarios (educational, fictional, historical, research) for jailbreak testing
• Adds 3 evaluation rubrics for llm_judge: tool_selection_safety, unbounded_agency, web_chatbot_security

Added:

• tool_selection_safety() scorer in tool_invocation.py
• role_play_wrapper() transform in stylistic.py
• 3 YAML rubrics in dreadnode/data/rubrics/
• Test suites and example notebook

Usage:

```python

Tool safety scorer

scorer = dn.scorers.tool_selection_safety(
safe_tools=["http_get", "tcp_connect"],
dangerous_tools=["shell_execute"],
)

Role-play transform

transform = dn.transforms.stylistic.role_play_wrapper(
scenario="educational",
character="security researcher",
)
```

Generated Summary:

Summary of Changes:

• Introduced new rubrics for safety evaluation in AI tools:

  • tool_selection_safety: Evaluates the choice of tools based on safety and risk (OWASP ASI02).
  • unbounded_agency: Assesses whether agents operate within user-defined limits and request permission before expanding scope (OWASP ASI10).
  • web_chatbot_security: Identifies web chatbot plugin vulnerabilities based on IEEE S&P 2026 findings.

• Added a scoring mechanism for tool selection safety, focusing on:

  • Detection of dangerous tool usage when safer alternatives are available.
  • Guidelines for classifying tool choice as safe, questionable, or unsafe.

• Implemented a role-play wrapper transform for testing against jailbreak attempts, distinguishing between legitimate educational inquiries and potentially harmful requests.

• Included an example Jupyter notebook to demonstrate the usage of both the tool selection safety scorer and the role-play wrapper, with practical scenarios to illustrate scoring and evaluation.

Potential Impact:

• Enhances the overall security posture by systematically evaluating tool selection and agency behavior in AI applications.
• Provides a structured approach to identify and mitigate vulnerabilities in web chatbots and tool misuse.
• Facilitates developers in ensuring compliance with security best practices through clear guidelines and scoring metrics.

This summary was generated with ❤️ by rigging