ArchestackBest of Open-Source AI: List on GitHub ↗ (opens in a new tab)

#1919th of 22 in Security

purplellama

Meta's safety classifiers and cybersecurity benchmarks for LLM inputs, outputs and code

Documentation ↗ (opens in a new tab)Website ↗ (opens in a new tab)Repository on GitHub ↗ (opens in a new tab)
Stars
4.5k
License
custom license
Last commit
Oct 2026
Language
Python

Missed the latest GitHub check; these facts may be older.

Overview

Purple Llama is an umbrella repo from Meta with input/output safeguards and security evals for generative AI. It includes Llama Guard 3 moderation models, Prompt Guard for prompt injection and jailbreak detection, and Code Shield for filtering insecure LLM-generated code. The CyberSec Eval benchmarks (v1 to v3) measure insecure code suggestions, code interpreter abuse, prompt injection, spear phishing and autonomous offensive operations.

Who it is for: Teams adding guardrails to LLM apps and researchers benchmarking LLM cyber risk

Strengths

  • Covers both defence (Llama Guard, Prompt Guard, Code Shield) and attack-side evals in one repo
  • CyberSec Eval maps to CWE and MITRE ATT&CK; public leaderboard on Hugging Face
  • Llama Guard 3 supports MLCommons hazard taxonomy, 128k context, and 1B, 8B and 11B-vision sizes
  • Code Shield example notebook provided; evals and Code Shield are MIT licensed

Weaknesses

  • Mixed licensing: MIT for evals and Code Shield, Llama Community licenses for models
  • README has no install, hardware or usage instructions; points to llama-recipes instead
  • Llama Guard weights are hosted on Hugging Face, not in this repo
  • Safeguard models are tied to the Llama family and its license terms

What it needs

  • Models: Llama Guard 3-8B, Llama Guard 3-1B, Llama Guard 3-11B-vision, Llama Guard 2, Llama Guard

Also in Security

See all 22
Also in Security
RankProjectScore
1shannonAutonomous pentester that exploits web apps and APIs using source analysis51.8k stars+3,117 (+6.4%) in 30 daysAGPL-3.092 out of 100
2ifixaiPython CLI that audits AI agents for policy, access and reliability failures45k starsApache-2.087 out of 100
3StrixAutonomous AI pentesting agents that validate vulnerabilities with working exploits75.1k starsApache-2.084 out of 100
#4ai-infra-guardRed-teaming platform for AI infra CVEs, MCP servers, agent skills and jailbreaks7.5k stars+669 (+9.8%) in 30 daysApache-2.079 out of 100
#5pentagiAutonomous multi-agent penetration testing system with sandboxed security tools34k stars+6,715 (+25%) in 30 daysMIT78 out of 100