Offensive Security Agent Swarm — An AI-powered multi-agent offensive cybersecurity platform built on OpenCode.
10 AI agents. 20 domain skills. 17 security tools. Attack chain synthesis. Persistent memory.
git clone https://github.com/kartikus/offseccode.git
cd offseccode
./install.sh| Component | Count |
|---|---|
| AI Agents | 10 (2 primary + 7 subagents + 1 advisor) |
| Domain Skills | 20 (19 security domains) |
| Security Tools | 17 (Kali Linux integration) |
| Knowledge Base | 2,027 atomic documents |
offsec-plan (blue) → plans the engagement
│
offsec-exec (red) → executes the assessment
├── recon (OSINT, subdomain, port scanning)
├── network (protocol exploitation, AD)
├── webapp (SQLi, XSS, SSRF, API)
├── exploit (vulnerability exploitation)
├── post-exploit (lateral movement, privesc)
├── cloudsec (AWS/Azure/GCP, K8s, CI/CD)
├── reporter (pentest reports, MITRE, CVSS)
└── advisor (second-opinion QA, Opus 4.8)
- Attack Chain Synthesis: Automatically chains individual findings into complete compromise paths (SSRF → internal network → RCE → root)
- Persistent Memory: Attack graph + session state survive restarts, crashes, and multi-day pauses
- Two-Model Verification: Independent advisor on different model lineage catches blind spots
- Domain-Aware KB: 2,027-document knowledge base with 50x search precision
- Human-in-the-Loop: Approval required at all critical exploitation gates
- Model Flexibility:
/model openai/gpt-5switches all agents instantly
- Kali Linux (or Linux with security tools)
- OpenCode installed
- Bun or Node.js
- DeepSeek API key (or any LLM provider via
/model) - Knowledge base at
/root/knowledge-base/kb-normalized/
- Run
opencode - Tab to
offsec-plan(blue) → plan your engagement - Tab to
offsec-exec(red) → execute the assessment - Or just describe your target to begin
/model openai/gpt-5— switch all agents to a different LLMAsk advisor to review the findings— get second opinionGenerate the final report— produce client-ready report
MIT