Build an automated red teaming framework for evaluating AI model safety, including jailbreak testing, bias detection, and harmful output assessment.
You are an AI red team lead specializing in LLM safety evaluation. Build an automated red teaming framework. ## Configuration - **Target Model**: [MODEL e.g. GPT-5 / Claude 4 / Llama 4 / custom fine-tuned model] - **Evaluation Focus**: [FOCUS e.g. safety / bias / toxicity / privacy / all] - **Application Context**: [CONTEXT e.g. customer support / code generation / medical advice / general assistant] - **Language**: [LANGUAGE e.g. Python] - **Budget**: [BUDGET e.g. 1000 / 5000 / unlimited API calls] ## Framework Components ### 1. Attack Library - Jailbreak prompt collection (200+ techniques) - Prompt injection variations (direct, indirect, multi-turn) - Social engineering attack patterns - Multi-language and encoding bypass attempts - Context manipulation and role-play exploitation - Adversarial suffix generation using GCG algorithm ### 2. Evaluation Dimensions - Harmful content generation (violence, illegal activities, self-harm) - Bias and stereotyping across demographics - Privacy leakage (training data extraction, PII generation) - Hallucination rate and factual accuracy - Policy compliance for [CONTEXT] use case - Overrefusal rate (legitimate requests incorrectly blocked) ### 3. Automated Testing Pipeline - Batch attack execution with retry and rotation - Response classification using judge LLM - ASR (Attack Success Rate) calculation per category - Statistical significance testing - Cost tracking against [BUDGET] ### 4. Reporting - Safety scorecard with pass/fail per category - Detailed failure analysis with example outputs - Comparison with industry benchmarks - Remediation recommendations ranked by severity - Regression test suite for fixed vulnerabilities Generate complete [LANGUAGE] framework with CLI, configuration management, and CI integration for continuous safety testing.
Free to copy and use. Compatible with Claude 4 Opus, GPT-5, Gemini 2.5 Pro.
Specify the target model and evaluation focus. Start with the automated jailbreak tests, then expand to bias and privacy evaluations. Review all results manually before sharing reports.
Initial release
Sign in and download this prompt to leave a review.