PurpleAILAB's Decepticon Hits 98% on Hacking Benchmarks With 16 Autonomous Agents
Decepticon is an open-source multi-agent red team framework that autonomously runs full attack chains inside a hardened Kali sandbox, scoring 98.08% on the XBOW benchmark
Takeaways
- What it is: Decepticon is an open-source autonomous red team agent with 16 specialist sub-agents covering the full attack kill chain.
- Benchmark: Scores 98.08% (102/104) on the XBOW validation benchmark — the highest published pass rate of any agent, open-source or commercial.
- Architecture: Two-network Docker design with a hardened Kali sandbox; agents use persistent tmux sessions to handle interactive offensive tools like msfconsole and evil-winrm.
- Governance built-in: Before any action, a Planner agent generates full RoE, ConOps, and an OPPLAN with MITRE ATT&CK mapping; out-of-scope actions are blocked at runtime.
- Flexible models: Supports Anthropic, OpenAI, Gemini, DeepSeek, Ollama, and more; subscription OAuth supported for Claude Max, ChatGPT Pro, and others.
- Free and open: Apache-2.0 licensed; install with a single curl command; also available as a pip SDK for building custom orchestrators on top.
Decepticon, from PurpleAILAB, is an open-source autonomous red team agent that does a lot more than scan ports and write a report. It runs full adversarial kill chains , recon, exploitation, privilege escalation, lateral movement, and C2 , the way a real attacker would, coordinated across 16 specialist sub-agents inside a hardened Kali Linux sandbox. The project has crossed 4,700 GitHub stars and reached #4 on GitHub Trending, and it's one of the most technically complete open-source offensive AI frameworks published to date.
The problem with "AI pentest tools"
Most tools in this space wrap a single LLM around a few shell commands and call it an AI hacker. The "AI + hacking" space is full of demos that run nmap and print a report , and that's not what Decepticon is. The real gap is that offensive security requires adaptive, stateful reasoning across multiple phases: you can't just prompt an LLM and expect it to chain a recon finding into a working exploit and then pivot laterally. Decepticon's architecture is built specifically to close that gap.
16 agents, one kill chain
Decepticon ships 16 specialist agents organized by kill chain phase. Each agent starts with a fresh context window per objective , no accumulated noise, no context degradation. Findings persist to disk and the knowledge graph, not agent memory. This "fresh context" model is a key design decision: rather than one giant LLM session that accumulates thousands of tokens of history and eventually degrades, each specialist gets only what it needs for its specific objective.
The agent roster covers the full attack surface:
- Soundwave (Planner): Interviews the operator and generates the Rules of Engagement, ConOps, Deconfliction Plan, and OPPLAN before any action is taken.
- Recon: Port scanning, service enumeration, DNS, subdomain discovery, and OSINT. Populates a Neo4j knowledge graph with hosts and services.
- Scanner / Detector / Verifier: A five-stage vulnerability research pipeline (Scanner → Detector → Verifier → Exploiter → Patcher) where the Verifier confirms findings using two independent methods before escalating anything Critical or High.