FEAT add crescendo persona scenario techniques (movie director, history lecture, journalist interview)

[precognitivem0nk]

What

Adds three persona-driven YAML variants of the single_turn_crescendo technique landed in #1665 (movie director, history lecture, journalist interview), plus a new pyrit/setup/initializers/components/scenarios.py initializer (ScenarioTechniqueInitializer, registry name scenario_technique, execution_order 3) that registers them.

Why

Following @rlundeen2's guidance on the closed PR #1657: ship YAML variants on top of #1665 instead of a new attack primitive. The scenario-techniques pattern is more flexible and lets dataset diversity drive crescendo coverage without new code paths.

Refs #388.

Files

• pyrit/datasets/executors/red_teaming/crescendo_movie_director.yaml (new, the persona @rlundeen2 specifically suggested)
• pyrit/datasets/executors/red_teaming/crescendo_history_lecture.yaml (new)
• pyrit/datasets/executors/red_teaming/crescendo_journalist_interview.yaml (new)
• pyrit/setup/initializers/components/scenarios.py (new, slots into the existing target/scorer/scenario_technique 1-2-3 chain)
• pyrit/setup/initializers/components/__init__.py (export)
• pyrit/setup/initializers/__init__.py (re-export)
• tests/unit/setup/test_scenarios_initializer.py (new, 26 tests covering YAML loading, schema, persona distinctness, registration, and the get_default_adversarial_target fallback)

Schema

Each persona YAML uses the same schema as crescendo_simulated.yaml from #1665: parameters [objective, max_turns], data_type: text, and a value block templated on {{ objective }} and {{ max_turns }}. num_turns defaults to 3, matching SeedSimulatedConversation's canonical default.

Opt-in by design

The personas live behind the new ScenarioTechniqueInitializer rather than in the core SCENARIO_TECHNIQUES list, matching @rlundeen2's note on #1657 ("I might put them in an additional scenario initializer vs the core"). Users opt in by including scenario_technique in their initializer config; out of the box, no Scenario auto-loads them. crescendo_simulated from #1665 remains the always-available default.

Verification

• pytest tests/unit/setup/: 207 passed
• pytest tests/unit/scenario/: 445 passed
• pytest tests/unit/: 7075 passed, 102 skipped (env-gated live targets), 0 failures
• ruff check, ruff format --check: clean on all new and modified files
• No em-dashes or en-dashes in any shipped YAML or source file (the test file has two intentional needles for absence assertions, with an inline comment explaining)
• Local rendering check on each YAML: loads via SeedPrompt.from_yaml_file, renders cleanly with render_template_value(objective=..., max_turns=3), no Jinja artifacts left
• Initializer registration check: running ScenarioTechniqueInitializer.initialize_async() in isolation registers all 3 personas plus the core specs

Notes

• DCO sign-off on the commit.
• Branch is currently behind upstream/main by 5 unrelated commits. Trial merge is clean and conflict-free. Happy to rebase if preferred.

[rlundeen2]

This looks great, ty @precognitivem0nk!

We're building some easier benchmarks soon, and I'm stoked to see how these all compare

[precognitivem0nk]

Thanks for the quick review and merge. Looking forward to the new benchmarks. Happy to add more persona variants when the comparison surfaces gaps.