BitterSecurity/Decepticon

Autonomous Hacking Agent for Red Team

Decepticon – Autonomous Red‑Team Agent

What it is – Decepticon is an open‑source, container‑based platform that lets a large‑language‑model‑driven agent run realistic offensive security engagements. It orchestrates a suite of specialist agents (recon, exploitation, post‑exploitation, etc.) that operate inside an isolated Kali Linux sandbox, generate full engagement documentation (ROE, OPPLAN, MITRE ATT&CK mapping), and interact with real tools such as nmap, msfconsole, Sliver, Ghidra, and BloodHound.

Why it matters – Most “AI‑hacker” demos stop at a single scan. Decepticon goes further: it follows multi‑step attack chains, maintains persistent interactive sessions, and isolates everything in a two‑network Docker setup (management plane + sandbox plane). The project is positioned as an “offensive vaccine” – using autonomous red‑team findings to improve defenses.


Key components

Component Role
Core stack (LiteLLM, PostgreSQL, Neo4j, LangGraph, Skillogy) Manages LLM routing, stores findings, runs the graph‑based knowledge base, and coordinates agent workflows.
Kali sandbox (Docker network sandbox‑net) Executes all commands/tools in a hardened, isolated environment separate from the management plane.
Specialist workloads (BloodHound CE, Sliver C2, Ghidra MCP, etc.) Spawned on demand via the orchestrator (ops_start(...)).
Web dashboard Optional UI launched from the CLI (/web).
Python SDK (pip install decepticon) Lets developers embed the agent factories, plugins, and safety gate into their own codebases.
Model profiles (eco, max, test) Tiered LLM selection with provider fallback (Anthropic, OpenAI, Gemini, Ollama, etc.).

Quick start (Docker‑compose)

# macOS / Linux / WSL2
curl -fsSL https://decepticon.red/install | bash
# Interactive onboarding – set provider API keys, choose a model profile
decepticon onboard
# Launch the full stack and drop into the terminal CLI
decepticon

Windows uses the same flow via PowerShell.

The CLI starts the always‑on management services; additional tools appear only when the orchestrator requests them. From the CLI you can open the optional web UI with /web.


Using Decepticon as a library

pip install decepticon               # core SDK
pip install "decepticon[neo4j]"    # add knowledge‑graph helpers

The SDK provides agent factories, a plugin bundle system, and a safety gate that forwards LLM calls and sandbox commands to the running services (configured via DECEPTICON_LLM__PROXY_URL and SANDBOX_URL).


Benchmarks & performance

Benchmark Difficulty Pass rate
XBOW validation (Level 1) Easy 100 % (45/45)
XBOW validation (Level 2) Medium 98 % (50/51)
XBOW validation (Level 3) Hard 87.5 % (7/8)
Overall All levels 98.08 % (102/104)

Full per‑challenge results, attack‑class matrix, and LangSmith traces are linked in the repo.


Typical use cases

  • Red‑team automation – Run end‑to‑end penetration tests without manually chaining tools.
  • Security research – Prototype new attack techniques by extending the agent roster or adding custom skills.
  • Defensive training – Generate realistic adversary emulation data for blue‑team exercises.
  • Product integration – Embed the SDK into SaaS platforms that need autonomous security assessments.

License & community

  • License: Apache‑2.0 (permissive, commercial‑friendly).
  • Support: Sponsorship via GitHub Sponsors, Discord community, live hosted app (app.decepticon.red).
  • Documentation: Extensive docs covering installation, architecture, agent roster, model configuration, and the “offensive vaccine” feedback loop.

Bottom line

Decepticon is a fully‑featured, production‑grade framework for autonomous red‑team operations. It combines LLM‑driven decision making with real security tooling, sandbox isolation, and a knowledge‑graph backend, making it one of the most complete open‑source AI‑pentest agents available today.

Related

  • Project
  • Project
  • Project
  • Project