Skip to content

AI red teaming and model testing

Attack your own AI apps on purpose to find jailbreaks, leaks and unsafe behaviour first.

5 productsUpdated 4 Nov 2026

AI red teaming tools send large sets of attack prompts at a model or application and report where it fails, for example through jailbreaks, prompt injection or data leakage. Security and engineering teams run them before launch and in CI. The best-known options are open source and free; commercial platforms add reporting, continuous testing and support. A clean scan does not prove safety, because attackers keep inventing new methods.

What to look for

  • Check which attack families are covered and how often the probe sets are updated.
  • Look for CI/CD support so tests run on every prompt or model change.
  • Confirm it can test your real endpoint (REST, agent, RAG app) and a bare model alike.
  • Ask how results are scored and whether you can add your own tests.
  • Compare open-source tools with paid platforms on reporting and compliance mapping.

AI red teaming and model testing products

Filter and sort in the directory

garak

NVIDIA

Open-source NVIDIA scanner that probes LLMs for jailbreaks, prompt injection, leakage and toxicity.

  • Self-hosted
  • Free (open source)
  • Free tier
  • Open source

Commercial platform for continuous AI red teaming and attack-led security of AI systems and agents.

  • Cloud (SaaS)
  • Price on request

Agent security platform with AI-SPM, runtime detection and response, access control and red teaming.

  • Cloud (SaaS)
  • Price on request

promptfoo

Promptfoo (now part of OpenAI)

Open-source CLI and library to evaluate and red-team LLM apps, agents and RAG systems in CI.

  • Self-hosted
  • Free (open source)
  • Free tier
  • Open source

PyRIT

Microsoft

Open-source Microsoft AI Red Team toolkit for scanning and attacking generative AI systems.

  • Self-hosted
  • Free (open source)
  • Free tier
  • Open source

Listed products are not ads and nobody paid to be here. Facts come from the vendor and public documents. A verdict appears only on products I have used myself.

More in AI security

By using this site, you agree to the Privacy Policy.