Send a request

Meta PurpleLlama (Llama Guard 4, Prompt Guard 2, LlamaFirewall)

by Meta

Free guard models for your own AI

Meta's open safety models and firewall that catch jailbreaks, prompt injection and insecure generated code.

Open SourceTrust 100For prosRuns offlinePython
Download

Checked sourceWhat does “checked” mean?github.com/meta-llama/PurpleLlamaMaker’s own licencechecked 6 hours ago

Have it set up for you instead:Request setup by OROBOOr start with the AI Compass: a written assessment at a fixed price of €490 (businesses only).

How it runs for you

Simulation starting …

Illustration, not a real test run

What you get

What you get

  1. 01Llama Guard 4 checks inputs and outputs for unsafe content
  2. 02Prompt Guard 2 detects jailbreaks and injection
  3. 03LlamaFirewall helps secure agents

For you if …

  • Developers self-hosting open models
  • Teams that want to run protection locally

Pick something else if …

  • If you have no model infrastructure of your own

How to start

In three steps.

What you need
  • Python
  • A GPU or enough compute
  • Acceptance of the Llama license
  1. 1Download the models on Hugging Face and accept the license
  2. 2Add them as a check before and after your model
  3. 3Set up LlamaFirewall for agents as the README shows

Safety check

7 of 7 checks passed

No known malicious patterns in the install instructionsactively maintainedMaker’s own licence

What does “checked” mean?

Trust 100
Who is behind it?
  • Taken by OROBO from the official source. The maker has not joined OROBO yet.
Is the code clean?
  • Source code publicly visible
  • Maker’s own licence
  • No known malicious patterns in the install instructions
Is it maintained?
  • Website reachable
  • Updated 6 days ago
Technical details & what we don’t check
  • Website loads fast
  • Website protection settings: 5 of 6 (website only, not the software)HSTS, CSP, nosniff, Frame-Schutz, Referrer-Policy

Claims about protection come from the maker. The trust score checks public signals, not effectiveness.

We automatically check origin, code patterns, licence, maintenance and reachability — again every day. We do not check whether the software fits your purpose, how it handles your data, or whether it works without errors. Last checked 6 hours ago. How OROBO checks

Alternatives for the same goal

ModelScanMake it safe · Protect AI (Palo Alto Networks)Check model files before you load themOpen SourceTrust high, 100Runs offlinegarakMake it safe · NVIDIAHave your language model attacked on purposeOpen SourceTrust high, 100For developersGuardrails AIMake it safe · Guardrails AI (Harvey)Check AI answers before they move onOpen SourceTrust high, 100For developersNeMo GuardrailsMake it safe · NVIDIAFirm guardrails for your chatbotOpen SourceTrust high, 100For developersGiskardMake it safe · GiskardScan LLM apps for hallucinations and leaksOpen SourceTrust high, 100For developersPyRITMake it safe · Microsoft (AI Red Team)Automate red teaming for generative AIOpen SourceTrust high, 100For developersSnyk Agent Scan (formerly Invariant MCP-Scan)Make it safe · Snyk (ex-Invariant Labs team)Scan installed MCP servers and skills for trapsOpen SourceTrust high, 100For developersCisco Skill ScannerMake it safe · Cisco AI DefenseCheck agent skills before you install themOpen SourceTrust high, 100For developers

Facts

AreaMake it safe
PriceOpen Source
LicenceMaker’s own licence
Runs onPython
Popular with developers★ 4,465 GitHub
Last updated30 Oct 2026
Sourcemeta-llama/PurpleLlama ↗
I am the maker