JoasASantos/NeuroSploit

Autonomous, multi-model penetration-testing harness — Rust, CLI-first.
by Joas A Santos & Red Team Leaders

NeuroSploit is an autonomous penetration-testing tool. You point it at a target; it does the security assessment on its own and hands you a report — recon, exploitation, chaining findings into real impact, and validation, with no human driving each step.

The key idea: the harness is the pentester, not the model. Instead of asking one LLM to "hack this", a Rust engine runs the methodology — it picks the right specialist agents for what it finds, runs them in parallel, re-sends every claim over the wire to confirm it's real, and has a second model check each finding before it's reported. The LLMs are the hands; the harness is the discipline that keeps them honest. That's why it finds more and invents less than running a model directly.

It's CLI-first (with a web console and a TUI), open source (MIT), and runs on your own API keys or a local subscription — nothing about your targets leaves your machine unless you send it.

Give NeuroSploit a URL, a source repo, a running app, or a host/IP. A Rust harness (tokio) drives a pool of LLMs — via API key or local subscription (Claude Code / Codex / Gemini / Grok) — and runs the whole engagement for you:

recon ─▶ pick only the agents that fit the surface ─▶ exploit in parallel
      ─▶ wire-check every claim ─▶ cross-model vote ─▶ chain into deeper impact ─▶ report

It won't invent findings: every claim needs a receipt (raw tool output or a file:line), is re-sent by the harness itself to confirm it reproduces, and is adjudicated by a second model before it reaches the report.

⭐ Useful? Star the repo. 🌐 Website & live docs: neurosploitai.github.io · Tutorial & Guide 📖 Deep dive: TUTORIAL.md · changelog: RELEASE.md · honest comparison: BENCHMARK.md

Linux / macOS (x64 & arm64):

curl -fsSL https://raw.githubusercontent.com/JoasASantos/NeuroSploit/main/setup.sh | bash

Windows (PowerShell):

irm https://raw.githubusercontent.com/JoasASantos/NeuroSploit/main/install.ps1 | iex

The installer sets up Rust if missing, builds the release binary, and links neurosploit onto your PATH. Build by hand instead:

git clone https://github.com/JoasASantos/NeuroSploit && cd NeuroSploit/neurosploit-rs
cargo build --release      # → target/release/neurosploit

# just run it — the interactive session asks everything, in plain language:
neurosploit

# black-box, subscription login (no API key):
neurosploit run http://testphp.vulnweb.com/ --subscription --model anthropic:claude-opus-4-8 -v

# white-box — review a source repo (file:line evidence):
neurosploit whitebox /path/to/repo --subscription --model anthropic:claude-opus-4-8 -v

# web console — a browser UI over the same CLI:
cd web && npm start     # → http://localhost:4173

No subscription? Use an API key for any provider instead — see TUTORIAL.md.

• The harness is the pentester. Agent selection is recon-driven; findings chain (reduce a foothold to a primitive — exec / read / write / SSRF / identity / secret — then pivot), each stage proven and non-destructive.
• Wire-bound evidence — every reported exchange is re-sent through a scope-guarded replay engine and checked against recorded traffic. A claim the wire contradicts can't stand as confirmed.
• Attack knowledge graph (Neo4j optional) — the run is a typed graph of hosts, services, findings, credentials and impact, spanning web and internal network. It directs the next move during the run and is written to graph/ even without a database.
• Deterministic validation — per-CWE validators judge the measured request/response bytes, not the model's prose. Grounding gate, POMDP belief and an anti-hallucination check demote anything the evidence doesn't support.
• Secrets never leak the report — credentials the agents observe become $SECRET_n in every human-facing report; raw values stay in the run's vault for re-testing only.
• Execution resilience — classified error handling (7 error types), exponential backoff (5 retries, 2s–90s), crash-state salvage (partial findings survive agent failure), agent lifecycle nudge (3 retries with simplified prompts), progressive budget warnings (70/85/95%), and run snapshot/resume so interrupted engagements pick up where they left off.
• LLM-powered dedup — deterministic fast-path for exact matches, plus a semantic LLM judge for findings with different wording but the same root cause.
• 486 markdown agents across web, code (SAST), AI, infra/AD/cloud, chains and recon — each a self-contained playbook you can read, edit or add to.
• Run via API key or subscription, 19 providers, including local & CPU-only (ollama: / llamacpp:) with no data leaving the host.

cd web && npm start     # → http://localhost:4173

A browser UI that drives the same CLI — no second engine. Launch and watch a run live, then per finding:

• Attack graph — host → service → finding → credential → impact, every edge marked confirmed / needs-validation / hypothesis, ordered by kill-chain stage.
• Wire evidence — the real request/response pairs behind a finding (baseline vs. attack, the identity pair, matrix probes, the oracle), each tagged with its wire provenance — the proof as bytes, not prose.
• Re-test one finding — re-run its proof through the replay engine; the dated history shows a finding drift to "gone" over time.
• Transcript viewer — any agent's run, split into navigable blocks; search spans every agent.
• Chains, dependency todolist, recorded traffic, deliverables, PDF export.

Every run is a self-contained folder, runs/ns-<ts>-<target>/:

GitHub / GitLab PR review gate (neurosploit pr <repo> <n> --fail-on critical), @neurosploit mention bot, one Jira card per finding, MCP server (neurosploit mcp), compliance mapping (PCI-DSS / HIPAA / SOC 2 / GDPR). See TUTORIAL.md.

For authorized testing only. Agents stay in scope, never run destructive / DoS actions, and require proof of exploitation. You are responsible for having permission for every target.

Joas A Santos & Red Team Leaders — MIT.

This is the slim, Rust-only distribution (neurosploit-rs/ + agents_md/). The older Python engine and web GUIs live on the v3.4.0 branch.