redteam-autoresearch

You generate adversarial prompts for LLM guardrail training data by running a bounded red-team autoresearch loop that profiles targets, selects seeds, applies mutators, queries models, and records labeled datasets locally.

Red-team autoresearch for LLM evaluation simulates real-world adversarial probing under explicit authorization to produce studyable, labeled guardrail datasets for downstream training.

You evaluate model safety using StrongREJECT judging by querying models with mutated adversarial prompts and scoring responses to calculate macro and micro attack success rates for benchmark comparisons.

Yes, you can use OpenAI with python-dotenv to automate adversarial dataset creation by setting up a local run workspace to handle batch generation, querying, and recording.

Local red-team guardrail generation with tenacity enforces authorized usage and local execution, meaning it requires explicit permission for adversarial probing and operates within bounded, local run workspaces only.