Feat/garak exploitation scenario

[Joel (aspire488)]

Description

Tests and Documentation

Summary

Adds a PyRIT-native implementation of Garak's exploitation probe family.

What's included

• "Exploitation" scenario with:
  • "JinjaTemplatePythonInjection"
  • "SQLInjectionEcho"
• Curated Garak-derived payload datasets and reusable echo template.
• Payload provenance and expected-output metadata.
• Payload-specific "SubStringScorer" instances.
• Stable payload-hash-based atomic attack names.
• Bounded payload execution through "max_payloads_per_technique".
• Generic baseline disabled because exploitation success is payload-specific.
• Focused unit tests covering:
  • Technique expansion
  • Seed-group construction
  • Payload metadata
  • Scorer selection
  • Stable attack naming
• Updated scanner documentation and notebook examples.

Safety / scope

The implementation tests whether a target model emits exploit material; it does not execute the generated Python or SQL against a runtime, database, or external system.

The initial datasets use a small non-destructive subset of the upstream Garak payload families.

Reference

Inspired by Garak's "exploitation" probe:

https://github.com/NVIDIA/garak/blob/main/garak/probes/exploitation.py

Closes #2534.

[Roman Lutz (romanlutz)]

Seems like a copy of #2576 which is covering the exploitation scenario. I'll close unless there's a reason for having this additionally.

[Joel (aspire488)]

Thanks for pointing that out — I hadn’t caught that #2576 was already covering the exploitation scenario.

That explains the overlap, so no issue with closing this one. I appreciate you flagging the existing work rather than letting a duplicate contribution move forward.

I’ll take the existing implementation in #2576 as the canonical direction and look for non-overlapping contribution opportunities in PyRIT going forward.