Strix — Overview
Part of STRIX — Map of Content in the STRIX Vault. Read this first.
In plain English
Imagine hiring a small team of penetration testers for an afternoon. You hand them one target -- a codebase, a website, an API -- and a scope document saying what they are and aren't allowed to touch. One of them (the "lead") reads the brief, splits the work into specialties (recon, injection attacks, access-control bugs, business logic), and assigns a tester to each. Each tester goes off, runs their own tools, tries their own attacks, and — when they find something real and can prove it — writes it up with evidence. When the lead is satisfied the team has covered the scope, they collect every write-up into one final report and hand it back to you.
Strix is that team, except every "tester" is an AI agent, and the whole engagement runs inside a disposable, sandboxed container instead of a real office. The "lead" is the root agent; the specialists are child agents; the container is a Docker sandbox pre-loaded with the same tools a human pentester would use (nmap, sqlmap, nuclei, a browser, an intercepting proxy). Nothing about the methodology is simulated -- the exploitation attempts are real HTTP requests, the proof-of-concept scripts really run, and the report is meant to be handed to engineering as-is.
Why this exists
Traditional penetration testing is slow (weeks of scheduling and manual work) and security scanners are fast but noisy (lots of false positives that still need a human to check). Strix's stated goal is to close that gap: autonomous testing with real proof-of-concept validation, so a finding in the report isn't a guess -- it's already been demonstrated to work. The project's own before/after framing for this bottleneck is "hours, not weeks."
Three design choices that make findings trustworthy
- Every finding needs a working proof-of-concept, not just a signature match. The agent has to actually demonstrate the vulnerability (a working exploit script, captured evidence) before it's allowed to file a report.
- Deduplication is judged, not just hashed. Before a new finding is accepted, an LLM specifically compares it against everything already found in the run and asks "is this the same root cause on the same asset," not just "does the text look similar" -- this catches near-duplicate reports that a simple string match would miss or a naive one would over-merge.
- The root agent never tests anything itself. Its system prompt explicitly forbids it from running scanners or exploits directly -- its only job is to read scope, decompose the target, spawn specialists, and aggregate. This keeps the "am I still in scope" reasoning concentrated in one place instead of repeated (and potentially drifting) across every agent.
The pipeline, at a glance
See STRIX — Pipeline Stages for the stage-by-stage table, and design/data-flow.md
for the full DFD.
Glossary
| Term | Plain meaning |
|---|---|
| Root agent | The one AI agent per scan that acts as orchestrator only — never runs tools directly, only spawns and aggregates. |
| Child agent | A specialist AI agent spawned by the root (or by another child) that does the actual hands-on testing. |
| Graph of Agents | Strix's term for the tree of root/child/grandchild agents and the messaging system that connects them (see STRIX — Architecture (HLD)). |
| Sandbox | The disposable Docker container (one per scan) all agents share, pre-loaded with pentest tools and a proxy. |
| Caido | An open-source intercepting HTTP proxy (like Burp Suite) that runs inside the sandbox and transparently captures every tool's outbound traffic. |
| Skill (internal) | A Markdown knowledge-pack file (strix/skills/...) an agent can load into its context — e.g. vulnerabilities/sql_injection.md, tooling/nuclei.md — distinct from the top-level skills/ folder, which holds instructions for external coding agents (Claude Code etc.) on how to drive Strix. |
| Lifecycle tool | A tool call that ends an agent's turn loop — finish_scan (root only), agent_finish (children), respond_to_user/wait_for_agents (interactive mode). |
| Scan mode | quick, standard, or deep — controls how thorough (and how long/expensive) a scan is. |
| CVSS | Common Vulnerability Scoring System — the standard 0-10 severity score every finding is rated with. |
| SARIF | Static Analysis Results Interchange Format — a JSON standard CI/CD tools (GitHub code scanning, etc.) understand, so Strix findings can gate a pull request automatically. |
| MTTA-style bottleneck | The manual-triage delay Strix targets: with proof-of-concept validation baked in, a human doesn't have to re-verify every finding before acting on it. |
Inputs and outputs
In: a target — local source directory, GitHub repo URL, live URL/domain, or an API contract (OpenAPI/Swagger file or Postman collection) paired with a live base URL — plus optional free-text instructions, a scan mode, and a budget/turn ceiling.
Out: everything lands in strix_runs/<run-name>/: penetration_test_report.md
(the executive narrative), one vulnerabilities/<id>.md per finding,
vulnerabilities.csv and vulnerabilities.json (machine-readable indexes),
findings.sarif (CI/CD-ready), and run.json (run metadata, cost, status).
Commands
| Command | What it does |
|---|---|
strix --target <path\|url\|repo> |
Run a scan (interactive TUI by default). |
strix -n --target <...> |
Headless scan — no TUI, prints results and exits (-n = non-interactive). Exit code 2 if vulnerabilities were found. |
strix --resume <run-name> |
Resume an interrupted or budget-stopped scan from its saved state. |
strix view [<run-name>] |
Open a local, read-only web dashboard over a run's results. |
strix auth login chatgpt / status / logout |
Manage running on a ChatGPT subscription instead of a metered API key. |
See process/runbooks/run-a-detection-scan.md for the full, copy-paste-verified
walkthrough.
Related: STRIX — Architecture (HLD), STRIX — Pipeline Stages, STRIX — Backends.