Source: OWASP GenAI Red Teaming Guide v1.0

OWASP's January 2025 guide that splits GenAI red teaming into model, implementation, system/infrastructure and runtime (human and agentic) evaluation.

Contents

Summary

A methodology document for security teams, ML engineers and risk owners. Its main contribution is scoping: it argues that red teaming an LLM product must cover far more than the model, and gives a phase-by-phase blueprint with checklists. (Phase detail here is from the guide’s table of contents as reported by search summaries; the full PDF was not read.)

Key claims

  • Four evaluation areas: model evaluation (inherent weaknesses such as toxicity, bias, jailbreak susceptibility), implementation evaluation (system prompts, guardrails, filters), system evaluation (the wider environment: APIs, storage, integrations, supply chain; called “infrastructure assessment” on the landing page), and runtime / human and agentic evaluation (how users or external agents manipulate the live system).
  • Audience spans cybersecurity staff, AI/ML engineers, red-team practitioners, risk managers, CISOs and business leaders.

Feeds

LLM red teaming

Suggest changes →