Strix: an autonomous AI hacker to audit a vibe-coded app

The problem Strix solves

You built an app with Lovable, Cursor or Bolt. It works. But 45% of AI-generated code fails OWASP Top 10 security tests (Veracode, 2025), and a classic scanner just lists suspicious patterns without ever checking whether they are exploitable. Strix takes the opposite approach: it's an AI pentest agent that plays the attacker. It runs your code, probes your endpoints, and only reports a flaw once it has actually exploited it.

How it works

Where a scanner is a linear script, Strix orchestrates several specialized agents that collaborate like a red team: reconnaissance, authentication, injection, SSRF, XSS, business logic. A lead agent coordinates while each sub-agent digs into its own area.

The cycle is a Think-Plan-Act-Observe loop: the agent thinks, plans, acts, observes the result, then adjusts its strategy. Not a fixed sweep, an investigation that adapts in real time. To act, it has concrete tools:

• HTTP proxy - it manipulates requests and responses to test protections
• Driven browser - it explores client-side paths, including XSS and CSRF
• Terminal - it opens sessions to test commands
• Python environment - it writes its own exploits to prove a flaw

The output is not one more alert: it's a proof of exploitation (PoC). Strix covers broken access control, SQL/NoSQL injection, SSRF, XXE, XSS, CSRF, JWT issues, race conditions, cloud misconfigurations - well beyond the OWASP Top 10.

Installation

You need Docker running (Strix works inside an isolated sandbox) and an API key from an LLM provider.

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

export STRIX_LLM="openai/gpt-5.4"
export LLM_API_KEY="your-api-key"

Then you point the agent at a target. The most relevant mode is scanning your own local code:

strix --target ./my-app
strix --target https://github.com/org/repo
strix -n --target https://my-app.com   # CI mode, no UI

Results are viewed with strix view. Open source, Apache 2.0 license, over 55,000 stars on GitHub. A SaaS version and an Enterprise offer exist alongside it, but the core runs on your machine, in CI, or even offline with a local model.

What to know before using it

Two things, learned from documented usage reports, before believing in a magic button:

• Code-aware mode is the right one - pointed at a black-box app with no source, Strix guesses API paths and burns its budget on 404s. On your local code, it maps precisely and goes deep.
• It consumes tokens - each loop sends a growing context back to the model. A quick scan on a small site can cost several dozen dollars in API. To be capped with strict limits.

That's exactly why a tool like this doesn't replace a human: it generates proven leads, but triage, prioritization and business context stay a matter of judgment. An agent with no one to read its findings is just expensive noise.