redteam-autoresearch

What problem does it solve?

This skill orchestrates a bounded red-team autoresearch loop to generate labeled guardrail training data for LLM safety. It enables attackers and judges to simulate real-world adversarial probing under explicit authorization, producing studyable datasets locally.

Core Features & Use Cases

• End-to-end autoresearch workflow: target profiling, seed research, batch generation, deterministic mutators, model querying, judging with StrongREJECT, recording, and archive-based novelty tracking.
• Data export for guardrails: prepares labeled prompts, responses, and metadata for training guardrail classifiers and detectors.
• Benchmark-ready: supports holdout seeds, difficulty strata, and macro/micro ASR reporting for model comparisons.

Quick Start

Set up a run workspace and start the bounded red-team autoresearch workflow to generate guardrail training data.