What is CyberAI?

CyberAI is a multi-agent orchestration layer for offensive security. Five specialized agents — Recon, Intel, Exploit, Report, Web3 — run a typed, auditable pipeline that turns a target into actionable attack paths and a validated report.

Two things set it apart from "LLM wrapper over nmap":

• OOB-driven exploitation. Blind vulns (SSRF, XXE, blind injection) are confirmed through out-of-band callbacks captured by phantom-grid, not guessed from response diffs.
• Agent-trust-aware design. Every banner and tool output is treated as untrusted input: sanitized, injection-scanned, and parsed before it ever reaches the LLM context. Adversarial thinking is a design input, not a disclaimer.

Reach beyond the network: the Web3 agent runs Slither static analysis and maps detectors to Immunefi severity tiers for smart-contract audits.

Quick Start

pip install cyberai

# dry-run: full pipeline, no real network calls
cyberai scan example.com --dry-run

# real scan with a local model (air-gapped, no cloud) and scope
cyberai scan app.target.com --provider ollama --scope "*.target.com"

cyberai status          # config and tool availability
cyberai replay <id>     # re-run a saved session

Trust-aware in one sentence: if Nmap reads a malicious SSH banner crafted to hijack the LLM context, the orchestrator neutralizes that vector before the data ever reaches the model.

Architecture

flowchart LR
    T([target]) --> O[Orchestrator<br/>typed · dry-run · budget]
    O --> R[Recon] --> I[Intel] --> E[Exploit] --> RP[Report] --> V([validated report])
    E <-->|inject ↔ correlate| PG[(phantom-grid<br/>OOB callbacks)]
    O --> W3[Web3 track<br/>Slither · Immunefi]

Trust boundary — injection-scan + banner sanitizer at every phase edge. Findings reach confidence = 1.0 only when confirmed out-of-band via phantom-grid.

Observability: SQLite audit log · session export/import · cyberai replay Interfaces: CLI · FastAPI dashboard (SSE) · MCP server (Claude Desktop)

Agents

Security design

• Agent trust boundaries — each agent runs with minimal permissions.
• Untrusted input handling — banners sanitized, length-capped, marked UNTRUSTED before LLM context.
• Prompt-injection detection — 33-pattern detector at every phase boundary; hits become MEDIUM findings, visible in the report.
• Scope enforcement — wildcard + !-exclusion matching honors HackerOne / Bugcrowd briefs (cyberai scope import).
• Audit trail — every agent action logged (JSONL or SQLite) with full inputs/outputs; sessions are replayable.

Quick start

git clone https://github.com/evkir/CyberAI.git
cd CyberAI
pip install -e .

cp config.example.yml config.yml
cp .env.example .env
# Edit .env — add OPENAI_API_KEY or ANTHROPIC_API_KEY (not needed for --dry-run)

# Dry-run: walks all 4 phases, no network, no API key
python -m cyberai scan example.com --dry-run

# Real scan, scope-restricted
python -m cyberai scan target.htb --scope '*.target.htb'

# Replay a saved session deterministically
python -m cyberai replay <session_id>

# Import a bug-bounty scope
python -m cyberai scope import h1 --program acme

# Status / config
python -m cyberai status

Web dashboard

uvicorn cyberai.web.app:app --reload
# http://127.0.0.1:8000  — session list, live SSE progress, report view

MCP server (Claude Desktop / Cursor)

python -m cyberai.mcp.server

Exposes recon/intel tools (nmap_scan, dns_enum, cve_search, epss_score, …) over the Model Context Protocol. See docs/mcp/integration.md.

Configuration

# config.yml
llm:
  provider: openai        # openai | anthropic
  model: gpt-4o
  max_tokens: 4096
  temperature: 0.2

phantom:
  grid_url: http://127.0.0.1:9090

output_dir: reports/
max_cost_usd: 0.0         # 0 = disabled; set to enforce a budget

Optional feature flags (default off, no-regression): use_native_tools, use_nuclei, use_llm_summary, use_judge.

Documentation

Related tools

Requirements

• Python 3.11+
• OpenAI or Anthropic API key (not required for --dry-run)
• Optional: phantom-grid (OOB), nuclei, slither, NVD API key

License

MIT — see LICENSE