An offensive AI swarm for authorized security testing.
Prometheus runs a multiagent system on top of LLMs and standard recon tools like subfinder, nuclei, and nmap. Each agent owns a role (recon, classifier, exploit, report) and writes back to a shared blackboard, so new findings automatically trigger follow-up work. Dry run on by default.
Use this only on systems you own or are explicitly authorized to test. Keep Dry run enabled unless your scope and permission are clear.
Architecture
Four layers, no cloud dependency. Go binary serves the API. Next.js shell proxies to it. Reports stay on disk.
┌────────────────────────────────────────────────┐
│ WIN95 SHELL (Next.js) /scan /swarm /findings │
└─────────┬──────────────────────────────────────┘
│ HTTP + WS (proxy /api/v1 → :8081)
┌─────────▼──────────────────────────────────────┐
│ GO API (Fiber) campaigns · scans · ws │
└─────────┬──────────────────────────────────────┘
│
┌──────┴───────┐
▼ ▼
CLASSIC SWARM (blackboard scheduler)
recon→classify recon ↔ classify ↔ confirm ↔ exploit
→plan→exec→rep ↺ adaptive: success seeds new recon
│ │
└──────┬───────┘
▼
LLM PROVIDERS · TOOL WRAPPERS · LOCAL MEMORY
claude/openai httpx/nuclei reports/
ollama/lmstudio nmap/+13 more campaigns/The blackboard, in five steps
Agents wake up when new facts appear, do their job, write the result back. Successful findings seed new recon. Failed exploits trigger an adapted plan pass.
TARGET_REGISTERED ↓ RECON → finds hosts, services, surface ↓ CLASSIFIER → buckets, drafts attack paths ↓ CONFIRM → validates the most promising ↓ EXPLOIT → scoped, dry-run-by-default ↓ REPORT → markdown + evidence chain ↺ ADAPTIVE LOOP → seeds follow-up recon
Watches the blackboard, picks the next move, allocates budget, routes prompts.
Subdomains, hosts, ports, content, JS endpoints. Scope-locked before any packet.
Buckets findings into vuln classes, drafts attack paths, ranks by likelihood × impact.
Builds PoC commands, runs them inside the safe-mode blocklist, confirms with evidence.
Word97-style markdown report: severity, evidence, remediation, repro steps. On disk.
Wrapped recon & exploit tools
Agents drive these directly. Wrappers enforce scope and emit structured output.
Built-in playbooks
Templated campaigns. prometheus playbook run <name>
Each terminal is a real agent in the swarm runner. The control strip locks target, scope, preset, dry run, and adaptive improvement before launch.
dashboard. Campaign overview, severities, per-agent readiness.
findings. Reports as papers on a desk. Word97-style reader.
Install in 60 seconds
# 1. Clone git clone https://github.com/NlCK01/prometheus-exe cd prometheus-exe # 2. Bootstrap config + install Go-based recon tools cp config.example.yaml config.yaml ./scripts/setup.sh # 3. Provide a model key export ANTHROPIC_API_KEY="sk-ant-..." # or OPENAI_API_KEY=... # 4. Build + serve API go build -o bin/prometheus ./cmd/prometheus ./bin/prometheus serve --port 8081 # 5. Start the shell (in another terminal) cd web && npm install && npm run dev -- --port 3001 # open http://localhost:3001
- Go 1.24+, Node 20+
- Anthropic, OpenAI, or local LLM (Ollama / LM Studio)
- Optional: Postgres for persistent state
Checks Go, Node, $PATH tools, config validity, DB.
Interactive setup. Provider, key, scope, writes config.yaml.
Skip recon-tool install for a leaner first run.
| command | does |
|---|---|
| serve | Start the API server (Fiber on :8081) |
| scan <target> | One-off scan against a target |
| campaign list | List all campaigns |
| campaign watch <id> | Live TUI dashboard, watch the swarm work |
| campaign stop <id> | Emergency-stop a running campaign |
| campaign explore <id> | Interactive attack-surface explorer |
| playbook list / run | List or run a playbook |
| ctf solve <target> | Autonomous CTF solver |
| ctf writeup <id> | Generate writeup from a CTF campaign |
| report <id> | Render the markdown report for a campaign |
| explain <cve|finding> | Plain-English explanation |
| doctor | Health check: deps, tools on $PATH, config, DB |
| init | Interactive setup wizard |
| config show / validate | Print or validate config (secrets redacted) |
| fp share | Export anonymized false-positive cache |
| mcp serve | Run MCP server for AI clients |
| demo | Scripted, network-free walkthrough |
- Swarm + classic engines
- 5 default agents
- 16 wrapped tools
- Win95 control panel
- Fine-tuned per-agent local models
- Container-scan agent
- Continuous monitoring mode
- Slack / Jira webhooks
- Multi-tenant red-team SOC
- Encrypted distributed blackboard
- Replayable campaigns
- Human-in-the-loop policy gates
Prometheus is built for red-team training, CTFs, and engagements with written authorization. Pointing it at systems you do not own or are not contracted to test is illegal in most jurisdictions.
Keep Dry run enabled until your scope is clear. Use Enforce exploit scope to block any command targeting outside your approved host and resolved IPs.
