Decepticon Emerges as Open-Source AI Red Team Agent for Autonomous Security Testing
Decepticon, a new open-source autonomous red team agent, integrates AI reasoning with hands-on security testing, connecting to models like Claude and OpenAI Codex for advanced attack path execution.
A novel open-source autonomous red team agent named Decepticon has been released, aiming to bridge the gap between artificial intelligence reasoning and practical security testing. This tool integrates with large language models such as Anthropic's Claude Code and OpenAI Codex, allowing security professionals to manage complex assessments directly from familiar coding environments. Unlike traditional security tools that focus on scanning and reporting, Decepticon is designed to autonomously navigate and exploit attack paths within predefined engagement parameters.
The Decepticon project, available under the Apache 2.0 license, supports a comprehensive range of red team operations. According to its GitHub repository, these capabilities include reconnaissance, exploitation, privilege escalation, lateral movement, and command and control. While these are stated capabilities, users are reminded that explicit written permission is required before testing any system, and the tool's effectiveness across diverse real-world networks is subject to ongoing validation.
At its core, Decepticon utilizes the Model Context Protocol (MCP) to connect external coding agents. This bridge provides operators with granular control, enabling them to inspect ongoing engagements, monitor progress, issue follow-up commands, check system states, and halt active operations. The actual red teaming tasks are executed within Decepticon's LangGraph server, with the coding assistant serving primarily as an interface for management and monitoring.
The agent architecture is built upon a foundation of specialized AI agents, with Decepticon listing 16 distinct agents. These agents cover a wide array of security domains, including discovery, exploitation, vulnerability research, Active Directory, cloud systems, smart contracts, and reverse engineering. To maintain focus and reduce historical noise, each objective is assigned a fresh context window, ensuring that the AI agent operates with relevant information.
Before commencing any testing, Decepticon meticulously generates rules of engagement, an operational plan, and coordination documents that are mapped to the MITRE ATT&CK framework. Commands are executed within persistent tmux sessions inside a Kali Linux sandbox environment, enhancing isolation and control. The system also features automatic prompt detection to ensure seamless interaction with tools like Metasploit, Sliver, and evil-winrm.
Decepticon's infrastructure is designed for robust operation, employing separate Docker networks to isolate management services from the sandbox and target systems. Findings are preserved across these boundaries using Neo4j. The standard setup requires Docker and Docker Compose v2, supporting macOS, Linux, and Windows. Specialist services like BloodHound CE and Sliver are initiated on demand.
The tool offers flexibility in model selection, supporting Anthropic, OpenAI, Gemini, and local Ollama instances, with configurable profiles to assign different model tiers to specific tasks. Developers report a high success rate of 98.08% on the XBOW validation benchmark, successfully completing 102 out of 104 challenges across easy, medium, and hard difficulty levels. However, it's important to note that these results were achieved in a specific lab setup with vulnerability category hints provided, and independent verification across enterprise networks is still pending.
Looking ahead, Decepticon has a planned feature called the "Offensive Vaccine loop," which aims to translate security findings into automated defensive changes and subsequent retesting. While currently a development goal, this feature underscores the project's ambition to create a more integrated and adaptive security testing lifecycle.