AgentSpyBoo

A Rust-based AI red team agent that runs a multi-step ReAct loop against a local OpenAI-compatible LLM server. Targets AMD Ryzen AI hardware (CPU-path today; NPU driver unblocked, inference runtime blocked — see PHASE-2-RECON.md).

Current state

• Branch main: Phase 2 CPU + Phase 2.5 refactor — multi-step ReAct loop, three chained tools (subfinder → httpx → nuclei), scope guards, severity-rated findings with dedup, structured Markdown reports, modular src/ layout.
• Phase 3 (next): org-level recon via Pius as a preflight step before the existing tool chain. Adds --org and --asn flags.

Architecture

• Loop: Observe → Think → Act. The LLM sees three tools, picks one per iteration, we run it, feed results back, repeat until the LLM decides to summarize.
• LLM backend: OpenAI-compatible chat completion API. Tested against Lemonade Server running Qwen3-1.7B-GGUF via llama.cpp on CPU. Any compatible server should work (Ollama, vLLM, etc.) but only Lemonade is exercised.
• Tool chain: subfinder → httpx → nuclei (curated templates, -severity medium,high,critical). Scope-guarded: every discovered host is checked against the --scope glob before being passed to the next tool.
• Findings: JSON files under findings/, severity-rated (info/low/medium/high/critical), deduplicated across iterations.
• Reports: Markdown under reports/, structured with Executive Summary, Findings Table, Methodology, Step Detail, Recommended Next Steps.
• Module layout (Phase 2.5 refactor): config.rs, scope.rs, llm/ (client, parser, prompt), tools/ (registry, locate, subfinder, httpx, nuclei), findings/ (models, parse), agent/ (react_loop, state), report/ (generator).

Build

cargo build --release

Dependencies: tokio, clap, serde, serde_json, reqwest (rustls-tls), anyhow, chrono. No SQLite, no tracing framework, minimal surface area.

Requires Rust 1.75+.

Runtime requirements

• A local LLM server exposing an OpenAI-compatible /v1/chat/completions endpoint. Default target is http://127.0.0.1:13305/api/v1 (Lemonade). Configurable via --base-url or the LEMONADE_BASE_URL env var.
• Recon tools on $PATH: subfinder, httpx, nuclei (with template set installed at the default location).

On the dev target these live at $HOME/go/bin after a standard ProjectDiscovery install.

Usage

agentspyboo recon example.com \
  --scope 'example.com,*.example.com' \
  --rate-limit 500 \
  --httpx-cap 150 \
  --max-iterations 5 \
  --verbose

Reports land in reports/<target>-<timestamp>.md, findings in findings/<target>-<timestamp>.json.

Hardware

Development target: GPD Pocket 4 — AMD Ryzen AI 9 HX 370 (Strix Point), Radeon 890M iGPU (gfx1150), XDNA 2 NPU (50 TOPS rated), 32 GB unified memory. ~770g handheld with built-in 2.5 GbE.

None of the hardware specifics are required for the current CPU path — the agent will run on anything with a working Rust toolchain and a reachable OpenAI-compatible LLM server. The Ryzen AI positioning matters for the deferred NPU roadmap, not today's behavior.

Honest state

• ✅ Phase 2 CPU runs end-to-end on the GPD against real targets (tested: hackerone.com, peterc3-dev.github.io, gitlab.com)
• ✅ Phase 2.5 refactor shipped — src/main.rs is 39 lines, all logic lives in typed modules
• ✅ NPU driver unblocked — patched amdxdna.ko loads on cold boot, xrt-smi + flm validate green. See PHASE-2-RECON.md.
• ⛔ NPU inference still blocked — FastFlowLM can't handle protocol-7 opcodes required by Qwen3/GGUF models. CPU path remains active until this unblocks.
• ⚠️ Nuclei on the target hardware CPU still times out on full template sweeps with large host lists; --nuclei-cap flag added to limit input hosts
• ⚠️ No automated tests yet. All verification has been manual end-to-end runs.

See RESEARCH.md for the research brief this project is based on, including references to comparable pentest agents (Shannon, PentestGPT, PentAGI, CAI), NPU backend options (FastFlowLM, ort crate + Vitis AI EP), and an honest gap analysis.