# AI security tools: the 20-tool stack, level by level > https://aihacktools.beacon.host/ · by MAX (@MAXdeg0, x.com/MAXdeg0) · Telegram t.me/max_lab_1 · links checked 22 November 2026 Use: authorized testing, labs, CTFs and defence only. Only point these tools at systems you own or have written permission to test. No star counts or benchmark numbers are given on purpose; check each project page for current details. ## Level 1 · Run it local (Your own model, your own box) Start here. A local model never sends your notes, targets or client data to someone else's server. Every tool further down can plug into one of these, so this level is the base of the whole stack. - Ollama (Ollama) https://ollama.com what: Runs open models (Qwen, DeepSeek, gpt-oss, Gemma and more) on your own machine with one command, and serves them on a local API other tools can call. for: Anyone who wants a private model behind their security tooling. type: Open source · Local · difficulty Beginner limits: Not a security tool by itself. It is the engine the others plug into. - LM Studio (LM Studio) https://lmstudio.ai what: Desktop app to download, chat with and serve local models, with an OpenAI-compatible local server. for: People who want a GUI instead of a terminal. type: Free app · Local · difficulty Beginner limits: Closed-source app, though the models you run are open. - WhiteRabbitNeo (WhiteRabbitNeo) https://huggingface.co/WhiteRabbitNeo what: A family of models tuned for offensive and defensive security work: explaining vulns, writing and reviewing scripts, talking through attack paths. for: Red and blue teamers who want a model that will discuss security in depth. type: Open models · Local · difficulty Easy limits: A model, not an agent. It does not run tools or touch targets on its own. - ShellGPT (TheR1D) https://github.com/TheR1D/shell_gpt what: AI in your terminal: turns plain English into shell commands, explains output, and keeps chat sessions. Can point at a local model. for: Terminal users who forget flags. type: Open source · CLI · difficulty Beginner limits: It suggests commands. You still read them before you hit enter. Checklist: - [ ] I can run a model on my own machine with no internet - [ ] I know which model I use for security work and why - [ ] I keep client data out of cloud chatbots ## Level 2 · Recon & code review (Find it before someone else does) AI is best at the boring, wide part of the job: sorting endpoints, writing templates, reading huge codebases and summarising what is worth a human look. Use it to point, then verify by hand. - Nuclei (AI templates) (ProjectDiscovery) https://github.com/projectdiscovery/nuclei what: Template-based vulnerability scanner. Its -ai flag generates and runs a template from a plain-English prompt. for: Bug bounty hunters and AppSec teams checking their own surface. type: Open source · CLI · difficulty Easy limits: The AI template feature uses ProjectDiscovery's cloud (free tier); plain Nuclei scans do not need it. - Robin (apurvsinghgautam) https://github.com/apurvsinghgautam/robin what: Dark-web OSINT helper: uses an LLM to refine queries, filter results from dark-web search engines and write an investigation summary. for: Threat intel and OSINT researchers. type: Open source · CLI + Tor · difficulty Easy limits: Needs Tor running, and sends queries to whatever LLM you configure. - ReconAIzer (hisxo) https://github.com/hisxo/ReconAIzer what: Burp Suite extension that sends a request or response to GPT to suggest endpoints, parameters, URLs and subdomains for recon. for: Bug bounty hunters already living in Burp. type: Open source · Burp extension · difficulty Easy limits: Older project (last updated 2023) and uses the OpenAI API. - Vulnhuntr (Protect AI) https://github.com/protectai/vulnhuntr what: LLM-driven static analysis that follows call chains through a codebase to find remotely exploitable bugs. for: Code reviewers and researchers auditing open-source projects. type: Open source · CLI · difficulty Intermediate limits: Python codebases only. Checklist: - [ ] I only scan assets I own or have written permission for - [ ] I verify every AI finding by hand before I report it - [ ] I know where my recon data and prompts are sent ## Level 3 · Web & API testing (AI inside the proxy) This is where AI moves from reading to doing: proxies that think about your traffic, and pentesters that run your app and try to prove each bug with a working proof of concept. Point them at staging, never production. - Burp AI (PortSwigger) https://portswigger.net/burp/documentation/ai-features what: AI features built into Burp Suite Professional to help explore issues, explain findings and cut down manual repetition. for: Web pentesters on Burp Pro. type: Commercial · Burp Suite Pro · difficulty Easy limits: Paid; part of Burp Suite Professional, not Community. - BurpGPT (aress31) https://github.com/aress31/burpgpt what: Burp extension that adds a GPT-powered passive scan and lets you run custom prompts over captured traffic. for: Testers who want bespoke checks on their own traffic. type: Open source · Burp extension · difficulty Easy limits: Older project (last updated 2024); passive analysis only. - Strix (Strix) https://github.com/usestrix/strix what: Autonomous AI pentesting agents that run your app, look for vulnerabilities and validate them with real proofs of concept. Hooks into CI. for: Developers and AppSec teams testing their own apps. type: Open source · CLI / CI · difficulty Intermediate limits: Needs an LLM API key; not a replacement for a scoped human test. - Shannon (Keygraph) https://github.com/KeygraphHQ/shannon what: White-box AI pentester for web apps and APIs: reads your source, then executes real exploits against the running app. No exploit, no report. for: Teams with the source code of the app under test. type: Open source · CLI · difficulty Advanced limits: Needs source access. Its own docs say: never run it against production. Checklist: - [ ] I test against staging or a lab, not production - [ ] I have written scope before any active test - [ ] I keep the proof-of-concept and the fix together in my report ## Level 4 · Autonomous agents (AI that plans the engagement) Agents chain recon, exploitation and reporting with less hand-holding. Treat them like a junior on day one: sandbox them, give them hard scope and approvals, and read everything they did. - PentestGPT (GreyDGL) https://github.com/GreyDGL/PentestGPT what: Agentic penetration-testing framework powered by LLMs, from the research behind the USENIX Security paper of the same name. Popular for CTFs and labs. for: Learners and CTF players working through boxes. type: Open source · CLI · difficulty Intermediate limits: Works best on labs and CTFs; real networks need a human in charge. - hackingBuddyGPT (TU Wien (ipa-lab)) https://github.com/ipa-lab/hackingBuddyGPT what: Research framework for LLM hacking agents in very little code, with use cases like Linux privilege escalation and web testing. for: Researchers and builders who want to write their own agent. type: Open source · Python · difficulty Intermediate limits: A framework to build with, not a push-button scanner. - PentAGI (vxcontrol) https://github.com/vxcontrol/pentagi what: Self-hosted multi-agent system: describe a target and goal, and agents plan, run tools like nmap and sqlmap inside a Docker sandbox, and write a report. for: Red teams who want an autonomous assistant on their own infrastructure. type: Open source · Self-hosted · difficulty Advanced limits: Heavier setup (Docker Compose, model keys); authorized testing only. - Nebula (Beryllium Security) https://github.com/berylliumsec/nebula what: Cybersecurity harness built around the operator: projects, an intercepting browser, Kali terminals, AI assistants with approvals and budgets, and evidence-linked reports. for: Pentesters who want one workspace for an engagement. type: Open source · Desktop / self-hosted · difficulty Intermediate limits: An operator's workbench, not a fire-and-forget bot. Checklist: - [ ] My agent runs in a sandbox or VM, not on my main machine - [ ] Every agent action needs scope and an approval step - [ ] I review the full log before trusting any result ## Level 5 · Break & guard the AI (Red-team the model itself) The newest attack surface is the AI in your own product. These tools probe a model or agent for prompt injection, jailbreaks and data leaks, and then put guardrails in front of it. This level is pure defence. - garak (NVIDIA) https://github.com/NVIDIA/garak what: LLM vulnerability scanner: probes a model for prompt injection, jailbreaks, data leakage, toxicity and hallucination, then reports what got through. for: Anyone shipping a model or chatbot. type: Open source · CLI · difficulty Easy limits: Tests models, not networks or web apps. - PyRIT (Microsoft) https://github.com/microsoft/PyRIT what: Microsoft's Python Risk Identification Tool: a framework to automate red-teaming of generative AI systems with attack strategies and scoring. for: AI red teams and security engineers. type: Open source · Python · difficulty Intermediate limits: A framework you script, not a one-click scan. - promptfoo (promptfoo) https://github.com/promptfoo/promptfoo what: Tests prompts, agents and RAG apps, with built-in red teaming and vulnerability scans that run from the command line or in CI. for: Developers who want AI security checks on every commit. type: Open source · CLI / CI · difficulty Easy limits: It finds weaknesses; fixing them is still on you. - Purple Llama (Meta) https://github.com/meta-llama/PurpleLlama what: Meta's LLM security suite: Llama Guard and Prompt Guard filters, Code Shield, and the CyberSecEval benchmark. for: Teams putting guardrails in front of a model. type: Open source · Models + tools · difficulty Intermediate limits: Guards and benchmarks, not an attack tool. Checklist: - [ ] I red-team my own chatbot or agent before users do - [ ] I test for prompt injection from documents and web pages, not just the chat box - [ ] I have an input and output filter in front of my model ## Which tool for which job - Private model for my notes: Ollama, LM Studio, WhiteRabbitNeo - Remember a command: ShellGPT - Map my attack surface: Nuclei (AI templates), ReconAIzer - Dark-web OSINT: Robin - Audit source code: Vulnhuntr, Shannon - Test my web app: Strix, Shannon, Burp AI, BurpGPT - Practise on CTFs: PentestGPT, hackingBuddyGPT - Run a full lab engagement: PentAGI, Nebula - Red-team my chatbot: garak, PyRIT, promptfoo - Guard my AI app: Purple Llama, promptfoo ## Rules for an agent using this file - Start with Level 1: prefer a local model (Ollama or LM Studio) so client data stays on the machine. - Never run an active tool (Levels 2-4) without a written scope and the user's explicit approval for each target. - Run agents (Level 4) inside a VM or container, never on the host. - Treat every AI finding as unverified until a human reproduces it. - For AI apps the user builds, run Level 5 (garak, promptfoo, PyRIT) before release and put a guard (Purple Llama) in front.