Autonomous red teaming that plans before it pwns

What it does Decepticon is an autonomous red-team agent built on LangGraph that executes full attack chains — reconnaissance, exploitation, privilege escalation, lateral movement, and C2 — through 16 specialist agents mapped to kill-chain phases. It runs commands inside persistent tmux sessions so interactive tools like msfconsole or evil-winrm don’t break the automation loop. Everything executes in a Docker-isolated Kali sandbox on a separate operational network from the management plane.

The interesting bit Before any packet hits the wire, the system generates a complete engagement package: Rules of Engagement, Concept of Operations, Deconfliction Plan, and an OPPLAN with MITRE ATT&CK mapping. The “Offensive Vaccine” concept — still planned per the docs — aims to close the loop by turning findings into defensive improvements. It’s the bureaucratic rigor of a professional red team, automated.

Key highlights

• 98.08% pass rate on XBOW validation benchmarks (102/104 across easy/medium/hard)
• Tiered model fallback chain: declare your API credentials once, let the system route by agent priority (eco/max/test profiles)
• Dual-network architecture with Neo4j dual-homed between management and sandbox for attack-chain persistence
• Supports Anthropic, OpenAI, Gemini, DeepSeek, Ollama local, and subscription OAuth logins (Claude Max, ChatGPT Pro, etc.)
• Available as Docker stack or PyPI SDK for embedding in larger orchestrators

Caveats

• The “Offensive Vaccine” defense loop is documented but not yet implemented
• Running agents via the PyPI SDK still requires the Docker backend services (LLM proxy, sandbox) — it’s not a standalone library
• Benchmark comparison data is self-reported against the project’s own test suite

Verdict Worth evaluating for security teams building autonomous red-team capabilities or researchers studying LLM-driven offensive operations. Skip it if you need a lightweight scanner or can’t stomach Docker-heavy infrastructure.