Meta launches Purple Llama for open model safety tooling

Meta launched Purple Llama, an umbrella project of safety and security tools for developers building on open generative AI models, named for combining “red team” attack testing with “blue team” defensive tooling. The initial release had two components: CyberSecEval, a set of benchmarks for measuring cybersecurity risk in language models — covering areas such as insecure code suggestions and the potential to assist offensive cyber operations — and Llama Guard, an openly released classifier model developers could use to filter risky inputs and outputs around their own deployments.

Meta framed the release as an answer to a problem specific to open-weight models: once weights are published, the developer cannot control every downstream deployment the way a closed API provider can, so safety tooling has to be something third parties can adopt themselves rather than a filter Meta operates centrally. The project launched alongside the AI Alliance, a coalition Meta helped form that included AMD, AWS, Google Cloud, Hugging Face, IBM, Intel, Microsoft, NVIDIA and others, positioning open safety tooling as a shared industry effort rather than a single company’s product.

Purple Llama arrived a few months after Meta’s Llama 2 release had made open-weight frontier-class models mainstream, and functioned as Meta’s answer to a recurring criticism of open release — that safety work is harder to enforce once weights are public. Llama Guard and CyberSecEval were subsequently updated in later versions and adopted as reference tooling by other open-model developers.