running-multiturn-attack-suite

What problem does it solve?

This Skill helps teams evaluate whether a stateful LLM application can be manipulated across multiple turns when a single prompt is not enough to trigger the behavior under test. It turns multi-turn red-team ideas into a disciplined, reproducible workflow with clear pass, block, partial, and inconclusive outcomes.

Core Features & Use Cases

• Turn-by-turn orchestration: Runs scripted conversations in order, including branching logic, state capture, and stop conditions.
• State-aware evaluation: Tracks observable signals such as returned text, tool calls, retrieval citations, persona drift, and summarization events.
• Outcome classification: Separates per-turn progress from per-script results and records the landing turn when an attack succeeds.
• Use cases: Assess chatbots with memory, agentic systems with persistent state, retrieval-backed assistants, and long-context apps where summarization or context overflow may change safety behavior.

Quick Start

Use the running-multiturn-attack-suite skill to analyze a signed engagement against a stateful chatbot with a multi-turn script corpus, per-script isolation, and observable state snapshots.