LlamaFirewall

lastmod 2025-07-14created 2025-07-141 edits

A framework for detecting/blocking AI security risks

About

• Purple Llama is a Meta-led umbrella project for an open-source AI trust and security ecosystem; it aims to bring together the various security tools, guides, evaluation tools (Eval), and benchmarks needed for generative AI (LLM) development and advance them with the community
• LlamaFirewall is a core component of the Purple Llama project; a framework that uses a multi-layer scanner architecture to detect and block in real time the various security threats (prompt injection, misalignment, code vulnerabilities, etc.) that can arise in generative AI (especially LLM-based chatbots/agents)
• Together with Purple Llama's other security/trust tools such as Llama Guard, Prompt Guard, Code Shield, and CyberSec Eval, it implements a comprehensive system-level AI safety layer
  • Llama Guard: a Llama 3-based model for detecting/blocking harmful input and output content. Can be included in LlamaFirewall policies and applied automatically at every input/output stage
  • Prompt Guard: specialized in detecting prompt injection/jailbreak attacks; integrated in LlamaFirewall as PromptGuardScanner
  • Code Shield: real-time detection of vulnerabilities and harmful code in LLM-generated code; integrated in LlamaFirewall as CodeShieldScanner
  • CyberSec Eval: Purple Llama's LLM cybersecurity benchmark tool. Used to test the actual security resilience and vulnerabilities of LlamaFirewall-based models
• Features low latency/high throughput, real-time application, transparent open source, and strong extensibility
• Lets security teams/AI developers quickly put safe guardrails in place for LLM-based chatbots, agents, and multi-step AI

See also

• llama

Favorite site

• PurpleLlama/LlamaFirewall at main · meta-llama/PurpleLlama
• LlamaFirewall - A framework for detecting/blocking AI security risks | GeekNews