# CyberStrike drops an AGPL harness for AI-driven red-teaming > Source: > Published: 2026-08-22 11:19:59+00:00 # CyberStrike drops an AGPL harness for AI-driven red-teaming What makes it different from the usual "ask GPT for a payload" scripts is the harness architecture. You don't just feed it a target and hope. The core loop runs: **recon → attack graph generation → tool orchestration → evidence collection → report**. Each phase is a pluggable module with a defined schema, so you can swap the LLM backend (local Llama-3-70B, [Claude](/en/tags/claude/), GPT-4o, whatever) without rewriting your exploit logic. **Key pieces worth knowing** **Attack graph DSL**— YAML-based, describes multi-step chains like "enumerate SMB → extract hashes → pass-the-hash → dump LSASS". The LLM expands high-level goals ("get domain admin") into concrete graphs at runtime.**Tool adapters**— First-class wrappers for nmap, bloodhound, crackmapexec, impacket, metasploit modules, and custom binaries. Adapters expose typed inputs/outputs so the planner can chain them reliably.**Memory layer**— SQLite-backed context store persists findings across runs. You can pause a campaign, switch models, resume — the graph state survives.**Safety rails**— Scope enforcement via CIDR/target allowlists, rate limiting per adapter, and a mandatory "dry-run" mode that logs planned actions without executing. The AGPL means any SaaS wrapper must expose these controls. **Getting a local instance running** ``` git clone https://github.com/cyberstrike/cyberstrike.git cd cyberstrike pip install -e .[local-llm] # pulls llama-cpp-python, FAISS, etc. cp config.example.yaml config.yaml # edit config.yaml — set your target scope, LLM endpoint, adapter paths cyberstrike init --workspace ./my-campaign cyberstrike run --goal "achieve domain admin" --dry-run ``` The dry-run output shows the generated attack graph with confidence scores per node. Once you're comfortable, drop `--dry-run` and it starts executing against the scope. **Where it shines and where it doesn't** **Strengths**: Handles multi-step logic that single-shot prompts butcher. The evidence collector auto-correlates logs, pcaps, and tool output into a timeline — huge for reporting. Local model support means air-gapped environments work.**Gaps**: BloodHound adapter only ingests JSON, doesn't drive the GUI. No built-in C2 framework integration yet (Cobalt Strike, Sliver, Havoc are on the roadmap). LLM hallucination on obscure protocol edges still happens — always verify before firing. **Licensing catch** AGPL-3.0 triggers if you expose CyberStrike as a network service. If you're building a commercial pentest platform on top, you must open-source your modifications. Several vendors have already reached out about dual-licensing; the maintainers seem open but haven't announced anything. Worth cloning if you run internal red-team exercises and want reproducible, auditable AI assistance. The codebase is clean — typed Python 3.11+, decent test coverage, and the module interfaces are stable enough to build custom adapters without fighting the core. [OpenAI spent months training models that were actively 14d ago](/en/news/5466/) [How a Hacker Used DeepSeek AI to Autonomously Attack Servers 21d ago](/en/news/4591/) [Anthropic AI Hacked 3 Orgs During Testing: A Deep Dive 21d ago](/en/news/4577/) [Three organizations got breached in a controlled exercise — and 22d ago](/en/news/4521/) [Next Retiring boomers are taking the institutional knowledge AI needs →](/en/news/7301/) [an AI side-hustle playbook](https://tanyan888.com/), with plenty of directly applicable cases.