FEAT: Add AdversarialPoetryConverter (verse reformulation jailbreak, arXiv:2511.15304) #4290

Is your feature request related to a problem? Please describe.

PyRIT has no converter for the adversarial poetry jailbreak. Bisconti et al. (arXiv:2511.15304, v3 Jan 2026) show that rewriting a harmful request as a poem is enough to get past refusal training on many current models, in a single turn and with no other trick. They converted 1,200 MLCommons AILuminate prompts into verse with a fixed meta-prompt and scored the responses with a 3-judge majority plus human review: 43.07% attack success across 25 models, against 8.08% for the same prompts in prose. Hand-written poems reached 62%. Results varied a lot by provider (DeepSeek 72.04%, Anthropic 5.24% on the converted set), which is exactly the kind of difference a red team wants to measure.

Describe the solution you'd like

An AdversarialPoetryConverter, built the same way as TenseConverter and ToneConverter:

• pyrit/converter/adversarial_poetry_converter.py: subclass of LLMGenericTextConverter, with converter_target and an optional prompt_template
• pyrit/datasets/converters/adversarial_poetry_converter.yaml: the meta-prompt. It asks for a short poem that keeps the request's intent in metaphor and imagery and ends with one plain line stating the request, which is the structure the paper describes
• Unit tests and a section in the converter docs

The paper does not publish its meta-prompt or its poems, so the template is our own and follows the paper's description. I'd say that in the docstring rather than claim we reproduce their numbers.

Before opening the PR I can run a small live check (prose vs poem on the same objectives, same target and scorer) and put the measured success rates in the PR description.

Scoping question

Should this start as a converter only, like TenseConverter and ToneConverter, or also be registered in the technique catalog?