Strix: An Open-Source Tool Where AI Does the Pen Testing — From Install to Scan

A rundown of the tool that blew up in the YouTube video "I Found The BEST New AI Hacking Tool." In the video it was called Stricks, but the actual name is Strix and the repo is usestrix/strix. GitHub stars have passed 60,000, and the license is Apache 2.0, so you can run it in your own environment at no cost.

Only use it on sites you own or client systems you've been authorized to assess. This tool actually sends requests to the target and attempts attacks. Unauthorized use is illegal in most countries.

What Strix actually does

Strix isn't a tool that just reads code and says "this looks risky." It spins up several autonomous AI agents at once, investigates the target and attempts attacks like a real hacker would, and then produces reproducible proof (PoC).

The implication is clear. Static analysis tools are full of false positives; Strix actually takes a hammer to it and shows you whether access is really blocked.

The agents split up roles and move simultaneously:

• Recon agent: maps hosts, tech stack, subdomains, and subdomain enumeration
• Exploit agent: attempts injection, auth bypass, privilege escalation
• Verification agent: actually reproduces what was found to check whether it's a false positive
• Knowledge sharing: agents share discoveries and chain attack paths together

The tooling is the same as what professionals use. Intercepting proxy (including a Playwright-based browser), shell execution, a Python exploit sandbox, and OSINT recon — with results tagged with CVSS scores and OWASP classification.

Why GLM-5.3 is the default recommended model

I checked why the YouTuber recommended GLM 5.3. The explanation was low refusal, and the official docs point the same way.

Per the official quickstart, the default recommended model is openrouter/z-ai/glm-5.3. The same doc lists other recommendations:

• Z.ai GLM-5.3 (OpenRouter) — the default pick
• OpenAI GPT-5.4
• Anthropic Claude Sonnet 4.6
• Google Gemini 3 Pro Preview
• DeepSeek V4 Pro
• Moonshot Kimi K3

One correction is worth making here. The video explained that models like Claude or GPT refuse security-testing prompts, but the official docs list Claude Sonnet 4.6 and GPT-5.4 as recommended models. Refusal tendencies do vary by model, and the fact that GLM-5.3 is adopted as the default is what the documentation guarantees. "Other models don't work" is the video's summary, and it differs from the official docs.

If you want to run it locally, point the LLM_API_BASE environment variable at an Ollama or LM Studio endpoint. The video recommended hardware like an RTX 5090 with a model like Cyber 27B, but that model name doesn't appear in the official docs. What the official docs guarantee is "connecting self-hosted models like Ollama and LM Studio."

Install (how to download)

Only two prerequisites: Docker has to be running, and you need one API key from a supported AI provider.

Step 1. Install

Use either method.

# official install script
curl -sSL https://strix.ai/install | bash

# or install with pipx
pipx install strix-agent

Step 2. Wire up a model

export STRIX_LLM="openrouter/z-ai/glm-5.3"
export LLM_API_KEY="your API key"

Settings are saved automatically to ~/.strix/cli-config.json, so you don't have to re-enter them every time.

Step 3. First scan

strix --target ./my-app-directory

On first run it automatically pulls the Docker sandbox image, and results are saved to strix_runs/<run name>.

How to specify targets

Targets can be specified five ways.

# 1. local codebase
strix --target ./app-directory

# 2. GitHub repo
strix --target https://github.com/org/repo

# 3. live web app
strix --target https://your-app.com

# 4. API spec + real server (full coverage without crawling endpoints)
strix --target ./openapi.yaml --target https://api.your-app.com

# 5. batch target list from a file
strix --target-list ./targets.txt

Pages that need auth can also be handled with an instruction.

strix --target https://your-app.com --instruction "Authenticate with: user:pass"

Three scan modes

The video introduced Quick, Standard, and Deep. Here's what the official docs say:

One thing to watch: Deep is the default. If you don't specify a mode in the command, it may run the slowest way. If you just want a quick check, it's better to pass --scan-mode quick.

The video said a Quick scan cost $2.64 on their target, but that amount varies by site and by how many tokens were used. Treat it as a reference figure.

How to look at the results

# open the most recent run
strix view

# open a specific run
strix view runname

A browser dashboard comes up where you can see the findings, a live map of the agent team, and past run history. The dashboard is bound to 127.0.0.1 by default and generates a link with a token in it. Don't send that link to anyone.

If you need to run it headless on a server without a display, use non-interactive mode.

strix -n --target https://your-app.com

When vulnerabilities are found it exits with a non-zero code, so you can attach it straight to CI.

Running it automatically on a server or in CI

It drops right into GitHub Actions.

name: strix-penetration-test

on:
  pull_request:

jobs:
  security-scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v6
        with:
          fetch-depth: 0

      - name: Install Strix
        run: curl -sSL https://strix.ai/install | bash

      - name: Run Strix
        env:
          STRIX_LLM: ${{ secrets.STRIX_LLM }}
          LLM_API_KEY: ${{ secrets.LLM_API_KEY }}
        run: strix -n -t ./ --scan-mode quick

It automatically narrows the scope to only the files changed in the PR.

Using it straight from a coding agent

You can also drop it into SKILL.md-compatible agents like Claude Code, Cursor, or Codex.

npx skills add usestrix/strix

That one command installs nine skills for running a pentest, fixing what's found, and CI scans. You can also connect your own MCP server and expose tools that way. Just put the list in ~/.strix/mcp-servers.json.

What does it cost

This is the part the video emphasized. In the past companies spent $5,000 to $10,000 on a single simulated attack; now you only pay for the tokens you use.

The video compared three options: API, GPU rental like RunPod, and local hardware. It said renting an 8x A100 cluster on RunPod runs about $15 an hour, and if you're uneasy about customer confidential data going out over an API, that approach is reasonable.

You can also run it on a ChatGPT plan you're already subscribed to.

strix auth login chatgpt
export STRIX_LLM="chatgpt/gpt-5.4"

That means it's processed within the subscription rather than costing per token used.

The video's test result

The YouTuber ran a Quick scan against their own static website. It was a small site of about 20 pages, and no immediately critical vulnerabilities were found. What they got instead was a report of short- and medium-term recommendations organized to OWASP standards.

The point isn't the result, it's the cost. The same level of review done by a person would take weeks; here it was a few dollars and a few minutes.

Rules you must follow

This tool actually sends requests to targets and attempts attacks. So don't use it outside these conditions:

1. It has to be a system you own directly
2. Or a client system where you have explicit written authorization
3. Test only within the contracted scope

The Strix project itself warns at the end of the README: "targets not listed above are illegal in most jurisdictions." Starting with your own site is safer.

Wrapping up

We've arrived at an era where AI goes past simply "reading code and telling you" and actually investigates, attacks, and produces evidence. Strix is a representative open-source tool for that shift.

Running it free in your own environment, Apache 2.0 so commercial use is free too, and agents collaborating to chain attack paths — that's the core of this tool.

To get started all you need is Docker and an API key. Just be sure to run your first scan on your own site. As long as you don't forget that it's a tool pointed at targets, it's a powerful one.

Reference video: I Found The BEST New AI Hacking Tool — this post was put together by cross-checking the video against the official docs (GitHub README, docs.strix.ai). Where the video and official docs differ, it's noted in the body.