Source: PyRIT (Microsoft) repository, docs and releases
Microsoft AI Red Team's open-source framework for composing targets, converters, attacks and scorers into automated, multi-turn red-team scenarios; v1.2.0 released 2026-11-13.
Contents
- What: open-source framework (MIT), README, documentation (1.1.0) and GitHub releases
- Publisher: Microsoft AI Red Team
- Date: v1.0.0 2026-07-24; v1.1.0 2026-09-04; v1.2.0 2026-11-13; accessed 2026-11-14
- Original: github.com/microsoft/PyRIT · docs · releases. The older
github.com/Azure/PyRITrepo was archived on 2026-03-27 and points here.
Summary
PyRIT (Python Risk Identification Tool) is a library for building automated red-team operations against generative AI systems. Where garak is a scanner of fixed probes, PyRIT is a toolkit: you compose a target, prompt converters, an attack strategy (single- or multi-turn, often driven by an attacker LLM) and scorers, and every exchange is persisted to memory for review.
Key claims
- Three interfaces: a Scanner CLI (
pyrit_scan,pyrit_shell) running built-in scenarios; CoPyRIT, a web GUI for human-led red teaming; and the Python framework. - Targets: OpenAI, Azure OpenAI/Foundry, Anthropic, Gemini/Vertex, Bedrock, Hugging Face, Ollama, Groq, OpenRouter, OpenAI-compatible APIs, HTTP, WebSocket, Playwright-driven web apps, and custom targets; v1.2.0 adds A2A, HackAPrompt and GitHub Copilot SDK targets.
- Scorers: true/false, Likert, classification and custom; backed by LLMs, Azure AI Content Safety or code. Since v1.1.0 a score can be “undetermined” (check
score.is_undetermined); v1.1.0 also added system-prompt-extraction, ShieldGemma and Roblox PII scorers. - Attacks named in the docs include Crescendo, TAP and Skeleton Key; releases add Best-of-N, split-payload, multilingual, system-prompt extraction, FigStep (v1.1.0) and GOAT plus context-wrap and character-split techniques (v1.2.0). The basic single-turn primitive is
PromptSendingAttack. - Converters transform payloads (v1.1.0 added ten, including bijection, character noise and Vigenère ciphers).
- Memory: conversations, scores and attack results stored in SQLite or Azure SQL (in-memory for tests).
- Scenario framework: standardized evaluations for content harms, psychosocial risks and data leakage, combining strategies and datasets.
- v1.0.0 set the v1 architecture (scenarios, attack techniques, executors, registries) with breaking changes; v1.2.0 adds persistent Operations and Findings for organizing engagements, and drops Python 3.10.