I’ve been following the space of AI-driven security testing since earlier this year, so when I came across Strix — an open-source AI penetration testing tool — I had to try it out. Unlike traditional scanners that blast out a list of potential issues, Strix runs autonomous AI agents against a target to find and validate vulnerabilities, with actual proofs-of-concept for anything it reports.

Setup

Getting started was straightforward:

  • Docker (running)
  • A GCP project with access to Vertex AI
  • Install via curl -sSL https://strix.ai/install | bash
  • pipx install "strix-agent[vertex]"

Configure the provider with environment variables:

export STRIX_LLM="vertex_ai/gemini-3.1-pro-preview"
export VERTEXAI_PROJECT="your-project-id"
export VERTEXAI_LOCATION="global"

More information on setup can be found here.

Then point it at a target for which you have authorization to scan:

strix --target https://your-app.com --scan-mode quick

The Run

I pointed it at a live web app I was testing. Strix launches a team of agents that work against the target, saving all results to strix_runs/<run-name> as it goes. I found the local dashboard (strix view) particularly useful — it runs entirely on your machine with no cloud upload, showing run status, findings, and a live map of what each agent is doing.

I would recommend first scanning in “quick” mode and then switching to “standard” mode once you get comfortable with the tool and the costs.

What Stood Out

Multi-agent orchestration. Strix doesn’t run a single prompt against your app. It spins up specialized agents for reconnaissance, exploitation, and validation that share findings with each other and chain vulnerabilities together — more like a small red team than a scanner.

Real exploit validation. This is the big one. Strix validates findings by actually running the exploit and producing a working proof-of-concept with reproduction steps. That’s a huge step up from static analysis tools, which are notorious for false positives.

Issues & Observations

A few notes from my experience:

  • Certain models like Gemini 3.6 Flash hit guardrails that can’t be bypassed — it refuses to scan for vulnerabilities. In my testing, Gemini 3.1 Pro Preview worked best.
  • A typical scan in “quick” mode costs between $10-$15 in tokens. It usually can be completed in under 20 minutes.
  • Babysitting can be required, as sometimes the tool wastes time on things that don’t matter.

Workflow

  1. Following up on scan results using an AI coding agent configured with an MCP such as Linear can be helpful for quickly documenting findings from strix_runs/<run-name>.
  2. Assign Linear issues to a background agent with context of the codebase to work on a respective fix.

Closing Thoughts

AI-native pentesting tools like Strix feel like a genuine shift in how application security testing gets done. Being able to run something like this locally, against your own app, and get validated findings with PoCs instead of a stack of false positives is compelling. I can see this fitting naturally into a CI/CD pipeline to catch issues before they reach production.

If blackbox-style testing can be completed by anyone with access to tools like Strix and an LLM subscription, you should run regular automated scanning for proper defense in depth in the AI-native world.

One important caveat: only run Strix against systems you own or have explicit written permission to test. It actively exploits whatever you point it at.