Purple Llama is Meta’s open umbrella project for building trust and safety into generative AI. It brings together benchmarks and safeguard models that help developers measure, and then reduce, the risks of working with large language models.
The name borrows from cybersecurity, where “purple teaming” combines offensive (red team) and defensive (blue team) work into one collaborative approach to risk. The project’s first components cover both sides of LLM safety. CyberSecEval is a benchmark suite that evaluates cybersecurity risks of LLMs, such as whether a model produces insecure code or assists with malicious activity. Llama Guard is a family of safeguard models that classify both user prompts and model responses against safety hazard categories, acting as an input/output filter for human-AI conversations. Code Shield adds inference-time filtering aimed at insecure code suggestions. The project is open source on GitHub with over 4,400 stars and was still actively updated in 2026; models are also published on Hugging Face. The evaluations and benchmarks are MIT-licensed, while the models use the Llama Community license, allowing research and commercial use.
Purple Llama is completely free and open source. The benchmarks are released under the MIT license and the models under the Llama Community license, with no paid tier or hosted service required.
AI teams shipping LLM-powered features who need credible safety evaluations, plus developers who want an open, self-hosted guardrail model instead of a proprietary moderation API.