meta-llama/PurpleLlama

What it is

Purple Llama is Meta's umbrella project for safety tooling around open generative AI models. It combines safeguards (Llama Guard 3, Prompt Guard, Code Shield) with the CyberSec Eval benchmark suite. Meta frames it as purple teaming: red-team evals plus blue-team defences.

Licensing split

The catalog summary calls the whole set MIT-licensed, but the page splits it. Evals and benchmarks are MIT. Any models, such as Llama Guard, fall under the corresponding Llama Community license. An engineer shipping a commercial product should read that license for the guard models before adopting them.

What each piece adds

• Llama Guard 3: input and output moderation models fine-tuned from Meta-Llama 3.1 and 3.2. They target the MLCommons standard hazards taxonomy. The 3.2-based release supports 7 new languages, a 128k context window and image reasoning. They are also tuned to flag helpful cyberattack responses and to stop malicious code from being executed in code-interpreter hosting environments.
• Prompt Guard: separates two attack classes. Prompt injection is untrusted third-party data in the context window that makes the model run unintended instructions. Jailbreaks are instructions meant to override the model's built-in safety features.
• Code Shield: filters insecure LLM code at inference time. The page lists three uses: insecure code suggestion mitigation, code interpreter abuse prevention and secure command execution. An example notebook is linked.

CyberSec Eval versions

• v1: insecure code suggestion frequency and compliance with malicious requests, based on CWE and MITRE ATT&CK.
• v2: adds code interpreter abuse, offensive cyber capability and prompt injection susceptibility.
• v3: adds visual prompt injection, spear phishing capability and autonomous offensive cyber operations.

A Hugging Face leaderboard is linked for comparing models.

What to do

Meta recommends checking all LLM inputs and outputs against content guidelines suited to the application, citing Llama 3's Responsible Use Guide. Getting-started resources for the safeguards are in the Llama-recipes GitHub repository, as part of a safety layer in the Llama reference system.

Caveats

The page gives no install commands, thresholds or benchmark scores. Meta's only stated result is that its initial evals found meaningful risks, both in recommending insecure code and in complying with malicious requests. Details are in the CyberSec Eval and Llama Guard papers.