root@tr0p:~#Home

ANUBIS

AI agents that run real offensive tools, prove the bug, and keep the receipts.

ANUBIS, branded 4NUB15, orchestrates specialised agents for recon, exploit, post-exploit, reporting and retest. They run real tools inside a Docker sandbox, record what they tried in a graph, and stay inside a scope you wrote down in a RoE file.

  • python3.12+
  • licenseApache-2.0
  • sandboxDocker · Kali
  • statuspre-alpha

InstallSource on GitHub

tr0p@kali: ~/anubis
$ anubis scan --target https://app.example.test --scope roe.yml --multi
[+] scope loaded: 1 client, 1 target, authorizer verified
planner  dispatch recon → surface of app.example.test
guard    nmap -sV app.example.test  allowed · in scope
guard    sqlmap … --os-shell  denied · destructive

# findings carry CVSS, OWASP and a PoC
[+] 1 finding(s): [high] SQL injection on /api/items
[+] report sysreptor saved to anubis_runs/reports/
$ 

Features

A scanner flags. ANUBIS tries to prove it, and refuses to touch anything outside the scope you gave it.

Specialised agents
An orchestrator plus recon, exploit, post-exploit, reporter and retester roles, each with only the tools its job needs.
Graph of agents
In --multi mode a planner dispatches workers in parallel. They exchange messages and write back to shared state.
Dual graph
An asset graph records what exists. An exploration graph records what was tried and why, with a coverage ledger, saved atomically per run.
Real tools in a sandbox
Commands run inside an ephemeral Docker container on a Kali image, wrapped in timeout. Nothing runs on your host.
Guardrails, not decoration
Scope and RoE with out_of_scope precedence, an approval rule engine, a noise governor, token and turn budgets, and nonce-marked untrusted output against prompt injection.
Skills as playbooks
Loadable SKILL.md packs for API recon, API security, OWASP Top 10, web app and general pentesting. The agent loads one when the task needs it.
Any LLM, with failover
A LiteLLM gateway picks the model, falls back across providers and enforces the budget. You bring your own key.
Reports you can hand over
One versioned contract, 4nub15.findings/v1, exported as JSON, CSV, Markdown, SARIF 2.1.0 or a reptor finding file for SysReptor.

Usage

Write the scope first. A scan without an authorised target does not start.

Scan a target self-hosted CLI

Pick a mode with `--scan-mode quick|balanced|deep`. `deep` turns on the multi-agent planner, and `--multi` does it for any mode. With `-n` the exit code is non-zero when findings exist, so it fits CI.

$ anubis scan -t https://app.example.test --scope roe.yml
$ anubis scan -t ./my-app --scan-mode deep
$ anubis scan -t https://app.example.test --scope roe.yml -n --multi

Export the report json · csv · md · sarif · sysreptor

Choose a format or let the file extension decide. Without an output path the file lands in `anubis_runs/reports/`.

$ anubis scan -t https://app.example.test --scope roe.yml --report-format sysreptor
$ anubis scan -t ./my-app --report-output findings.sarif
$ cat findings.sysreptor.json | reptor finding

Inspect skills · runs

List the skill packs the agents can load, or open a finished run in the viewer.

$ anubis skills
$ anubis view

Agent loop

Every command an agent wants to run goes through the same gates before it touches a target.

  1. Scope The RoE file names the client, targets and authorizer. Out of scope always wins.
  2. Plan The planner splits the goal into intents and dispatches workers.
  3. Guard Approval rules, noise limits and budgets allow or deny each command.
  4. Sandbox The command runs in the Kali container, under a timeout.
  5. Record Assets, attempts and findings go into the dual graph.
  6. Report Findings export with CVSS, OWASP, PoC and evidence IDs.

Numbers

MetricValue
Agent roles6
Skill packs5
Report formats5
Guard layers5

Install

Clone and install in editable mode. You get the anubis console command.

$ git clone https://github.com/marcostolosa/anubis.git && cd anubis && pip install -e ".[dev,sandbox,db]"
  • Requires Python 3.12+ and Docker for the sandbox.
  • Bring your own LLM key. LiteLLM supports most providers.
  • Pre-alpha: expect rough edges. The proxy recorder is not in the normal scan path yet.
  • Authorized targets only: your own lab, your organization, or a program within its published scope.