Test the security of AI apps and agents

Red-team your chatbot, scan an LLM for weak spots, or let an AI agent pentest your own app.

For teams that ship AI features or agents and want to find weak spots before attackers do. The list starts with tools that test your AI app, then AI agents that test your code and your servers.

8 picks in 3 sections · checked by an editor on Oct 30, 2026 · How we pick

01Test your AI app4 picks

  1. promptfoopromptfoo/promptfoo

    Start here. Red-team prompts, jailbreaks and data leaks from one config.

    CLI · Library / SDK · MIT · active 22 hours ago · ★ 27k

    93worth
  2. garakNVIDIA/garak

    Pick this to scan a model for many known attack types.

    CLI · Apache-2.0 · active 4 days ago · ★ 9.8k

    84worth
  3. iFixAiifixai-ai/iFixAi

    Pick this to audit what an AI agent does with its tools.

    CLI · Apache-2.0 · active yesterday · ★ 39k

    89worth
  4. GiskardGiskard-AI/giskard-oss

    Pick this to find unsafe or made-up answers before users do.

    Library / SDK · Apache-2.0 · active 3 days ago · ★ 5.9k

    74worth

02AI pentesters3 picks

  1. Strixusestrix/strix

    Pick this to let AI agents look for real bugs in your own app.

    Docker / self-host · CLI · Apache-2.0 · active 3 days ago · ★ 73k

    90worth
  2. ShannonKeygraphHQ/shannon

    Pick this to test web apps and APIs, with proof for each finding.

    Docker / self-host · CLI · AGPL-3.0 · active 5 days ago · ★ 51k

    89worth
  3. pentagivxcontrol/pentagi

    Pick this for a team of agents that runs a full pentest in Docker.

    Web · Docker / self-host · CLI · MIT · active last week · ★ 33k

    81worth

03AI code scanners1 pick

  1. Codex Securityopenai/codex-security

    Pick this to scan your code for security bugs with an AI model.

    CLI · Library / SDK · Apache-2.0 · active 23 hours ago · ★ 11k

    85worth

More in Security

Other groups

How we pick

An editor picks 6 to 12 projects for one job and writes when to choose each one. Every pick had a commit in the last 6 months. It is not archived. Its worth score is 60 or more. An editor breaks a rule only for a reason, and writes that reason in the pick's note. We check every list again each quarter. A pick goes when it breaks a rule or a clearly better project appears.