Decepticon

• Running an authorized internal red team exercise that chains initial access, privilege escalation and lateral movement through Sliver C2 sessions
• Mapping and exercising Active Directory attack paths with BloodHound CE in a lab or a scoped client environment
• Testing whether a blue team detects realistic, ATT&CK-mapped adversary behavior during a planned engagement
• Running a registered red-team, pentest, bug-bounty or vulnerability-research workflow against scope the operator has approved
• Following an engagement's transcript, OPPLAN and scope from Claude Code or Codex through the MCP server

Decepticon stands out among open-source AI security agents for aiming at red teaming rather than application scanning: it plans objectives, keeps interactive shells alive, and brings up C2 and BloodHound when the engagement calls for them. Its scope handling and network separation are documented in more detail than most projects in this space, which matters for a tool that can actually compromise hosts. The cost is operational weight, a multi-container stack, model spend and an expert operator, and its headline benchmark is self-reported and measured on web challenges, which says little about the post-exploitation work that sets it apart. Treat it as an assistant for experienced operators on authorized engagements, not a hands-off pentest service.

About Decepticon

Decepticon is built for red teaming rather than vulnerability scanning: instead of checking a target against a list, it pursues objectives through a chain of techniques the way an adversary would, from recon and initial access through privilege escalation, lateral movement and command and control. Every Red engagement starts with the operator entering in-scope targets and exclusions; the launcher normalizes them, shows the exact resulting scope for review, and writes a machine-readable rules-of-engagement file (with a seven-day authorized window when created from the CLI) that the runtime enforces on gated tool calls and refuses to run without. An optional Plan mode uses a planning agent, Soundwave, to discuss assumptions and draft a threat profile, concept of operations, deconfliction, abort and cleanup documents; it cannot execute offensive work or change scope. In execution, the Decepticon orchestrator owns an OPPLAN of dependent objectives and hands each one to a specialist with a fresh context window, so long engagements do not accumulate noise. Specialists run tools such as nmap, sqlmap, Impacket, msfconsole and the Sliver C2 client inside persistent tmux sessions in a Kali container on a separate operational Docker network from the management plane, findings go to a Neo4j knowledge graph, and heavyweight workloads such as BloodHound CE, Sliver C2 and Ghidra-based reversing are started on demand from a fixed allowlist. Models route through LiteLLM with eco, max and test profiles and automatic fallback across Anthropic, OpenAI, Google, DeepSeek, xAI, Mistral, OpenRouter, local Ollama and others, including subscription sign-in. You can drive it from its terminal CLI, an optional web dashboard, Claude Code or Codex through its MCP server, or build on it with the Python SDK. Bitter Security also runs a hosted workspace with Solo and Pro plans, but new purchases are currently closed and live access is by request. The tradeoffs: you operate a multi-container Docker stack and pay your model provider, it is a tool for experienced operators on systems they are explicitly authorized to attack, the project's headline XBOW benchmark result is self-reported, and parts of the vision, such as the Offensive Vaccine attack-defend-verify loop, are still planned.