PyRIT

Microsoft's AI Red Team built PyRIT to automate the tedious parts of probing generative AI systems for safety and security failures, and released it under the MIT license. The framework is modular: targets wrap the system under test, covering OpenAI, Azure, Anthropic, Google, and HuggingFace models plus custom HTTP endpoints, WebSockets, and even web apps driven through Playwright; converters transform payloads with encodings and obfuscations; scorers judge responses with true/false, Likert scale, and classifier-based checks; and attack executors run single-turn probes or multi-turn strategies such as Crescendo, TAP, and Skeleton Key. Results persist to SQLite or Azure SQL memory so campaigns are auditable and resumable, and the project ships a command line scanner, a CoPyRIT web GUI, and a programmatic Python API. Installation is pip install pyrit on a recent Python, and no GPU is needed since attacks run against external endpoints. Security teams use it to run repeatable jailbreak and prompt injection campaigns at a scale manual red teaming cannot match.