---

name: attacks-scenarios description: "Choose, configure, and interpret PyRIT attack executors, techniques, scenarios, benchmarks, workflows, and prompt generators safely." disable-model-invocation: true metadata: disco-role: operating license: MIT

PyRIT attacks and scenarios

Use this sub-skill when the task is to plan or implement a PyRIT attack run, scenario campaign, benchmark, prompt-generation workflow, or attack-technique selection. This skill owns orchestration semantics: which executor or scenario should drive the run, what pieces must be wired together, how retries/concurrency affect execution, and how to interpret attack/scenario results.

Safety boundary: examples and helper scripts here are no-secret and no-network by default. Do not send prompts to a live model or service unless the caller explicitly supplies an approved target configuration, credentials, scope, and data handling rules.

Route first

User need	Use this sub-skill for	Route elsewhere for
Pick an attack algorithm	Single-turn, multi-turn, compound, streaming, benchmark, workflow, prompt generator, scenario vs atomic attack decisions	Target/scorer constructor details in targets-scorers
Configure attack components	AttackConverterConfig, AttackScoringConfig, AttackAdversarialConfig, AttackParameters, memory_labels, retry and concurrency placement	Converter/dataset construction in converters-datasets
Run a scenario campaign	Technique selection, baseline policy, DatasetAttackConfiguration, initialize_async, run_async, scenario result interpretation	Initialization/memory backend setup in setup-memory-core
Use pyrit_scan or backend	Scenario semantics and what techniques/datasets mean	CLI invocation, backend process, and scanner flags in cli-backend-scanner
Debug attack/scenario failures	Ownership of branching, async/concurrency, missing defaults, optional heavy paths, memory/result lookup	Credential, target, and scorer failures in targets-scorers

Required first checks

1. Confirm PyRIT is initialized for the intended memory backend before executing anything that writes prompts, scores, attack results, or scenario results. Route setup details to setup-memory-core.
2. Identify the objective target: the system under test. Do not confuse it with an adversarial target, which PyRIT controls to generate attack prompts for adaptive attacks.
3. Decide whether the run is a one-objective attack, many-objective atomic attack, scenario campaign, benchmark, workflow, or prompt generator.
4. Decide whether scoring is needed. Without an objective scorer, many attacks can execute but report UNDETERMINED rather than success/failure.
5. Check modality compatibility: converters and next_message data types must produce inputs accepted by the objective target; multimodal scorers must be able to evaluate the returned modality.
6. Bound concurrency and retries before running against rate-limited targets or expensive adversarial/scorer models.

Fast workflow selection

• Need one direct prompt or a simple dry run: use PromptSendingAttack with a target and optional AttackScoringConfig.
• Need a fixed conversation script: use MultiPromptSendingAttack or prepended_conversation/next_message on a single-turn attack.
• Need adaptive back-and-forth: use RedTeamingAttack, CrescendoAttack, TreeOfAttacksWithPruningAttack/TAPAttack, or PAIRAttack with AttackAdversarialConfig.
• Need fallbacks under one objective: use SequentialAttack with SequentialChildAttack and a SequenceCompletionPolicy.
• Need a campaign across objectives and techniques: use a Scenario; configure datasets, techniques, baseline policy, concurrency, retries, and labels before initialize_async().
• Need systematic measurement: use benchmark executors such as question-answering or fairness/bias benchmarks.
• Need generated prompts for later attacks: use prompt generators; treat GCG and HuggingFace-backed paths as optional heavy dependencies.
• Need scanner commands: keep semantics here, then route command construction to cli-backend-scanner.

Bundled operating references

• references/attacks-scenarios-reference.md — orchestration model, executor/scenario selection, attack/scenario workflows, composition patterns, concurrency and result interpretation.
• references/api-reference.md — distilled import/signature tables and practical constructor notes for attacks, scenarios, techniques, benchmarks, prompt generators, and results.
• references/troubleshooting.md — diagnosis matrices for ownership mistakes, missing components/defaults, modality mismatches, async issues, rate limits, optional GCG/benchmark paths, and memory/result lookup.
• scripts/attack_scenario_smoke.py — no-secret/no-network smoke helper that imports core PyRIT attack/scenario classes and prints signature metadata without sending prompts.