Vasco0x4/AIDA

AIDA – AI‑Driven Autonomous Pentesting Agent

What it is – AIDA (AI‑Driven Security Assessment) is an open‑source tool that turns a large language model (Claude, Gemini, OpenAI Codex, etc.) into an autonomous penetration‑testing assistant. It runs inside a Docker container that ships with common security tools (nmap, sqlmap, ffuf, nuclei, …) and can install additional ones on demand. The LLM drives the whole workflow: it plans the assessment, issues commands to the container, writes custom Python payloads, crafts raw HTTP requests, and records every step in a persistent notebook that can be resumed later.

---

Core capabilities (as described in the README)

Capability	How it works
Execution environment	A pre‑built aida-pentest Docker image (~2 GB) containing a Linux shell and a suite of pentesting utilities. The agent can also attach the external Exegol container for a larger toolset.
Tool‑calling LLM	Any LLM that supports tool‑calling (Claude Code, OpenAI Codex, Kimi, Gemini, or any OpenAI‑compatible API) can be used. The model is invoked via python3 aida.py and receives the assessment scope, then issues calls such as execute(), scan(), http_request(), etc.
Dynamic Python payloads	The agent can generate and run ad‑hoc Python scripts inside the container for custom exploitation or data processing that off‑the‑shelf tools cannot handle.
Burp‑like HTTP control	Through the http_request() tool the model can send arbitrary HTTP requests, inject stored credentials via {{PLACEHOLDER}}, and manipulate headers, cookies, bodies, etc.
Persistent notebook & reporting	All findings, command outputs, and reasoning are logged. Findings are added with add_card() which automatically scores them with CVSS 4.0. The dashboard (http://localhost:31337) lets you view the attack timeline, export PDF reports, and filter across multiple assessments.
Authentication & notifications	v1.1.0 adds JWT‑based user roles, a first‑run wizard, and optional Telegram/Slack/Email alerts (with PDF attachments).

---

Quick start (from the README)

# Clone and start the UI container
git clone https://github.com/Vasco0x4/AIDA.git
cd AIDA
./start.sh               # launches the dashboard on http://localhost:31337

# Run the agent (defaults to Claude Code)
python3 aida.py --assessment "target-corp"

Optional flags let you pick a specific model (--cli codex), run without prompts (--yes), or expose the UI on your LAN/public domain.

---

Typical workflow

1. Define scope – tell AIDA the target URL, subdomains, and any exclusions.
2. Let the LLM plan & act – it calls scan(), subdomain_enum(), ssl_analysis(), etc., installing missing tools if needed.
3. Review findings – the notebook lists each vulnerability with:
   • the command that discovered it
   • raw output
   • the model’s reasoning
   • an auto‑generated CVSS 4.0 score
4. Export / hand‑off – one‑click PDF report or continue the engagement later.

---

Real‑world impact claimed

The README lists several CVEs (e.g., CVE‑2026‑49869, CVE‑2026‑50189) that were reportedly discovered by AIDA + Claude and have been published in public vulnerability databases.

---

Limitations & cautions (as noted by the authors)

• Model quality matters – deeper assessments require a more capable LLM; the tool is “model‑agnostic” but a weak model will produce shallow results.
• Container exposure – the dashboard should stay on localhost or a trusted LAN; the authors warn against exposing it publicly.
• Tool‑calling reliability – the agent relies on the LLM correctly invoking the predefined tool functions; malformed calls could stall the run.
• Legal/ethical responsibility – the user must have permission to test the target; AIDA only automates the grunt work, the human reviewer remains responsible for triage and disclosure.

---

License

AGPL‑v3 – the code must remain open source and any network‑served modifications must be shared.

---

Who might find this useful?

• Red‑team engineers looking for a rapid, AI‑assisted reconnaissance and exploitation scaffold.
• Bug‑bounty hunters who want an LLM to automate repetitive scanning while they focus on verification.
• Security researchers interested in evaluating how far current LLMs can drive end‑to‑end pentesting.
• Organizations that want an internal “assistant” to run baseline assessments on owned assets (with proper authorization).

---

Bottom line

AIDA is a genuine open‑source project that combines LLM tool‑calling with a ready‑made pentesting container to deliver an autonomous security‑assessment agent. It provides a web dashboard, persistent notebooks, CVSS scoring, and multi‑model support, all orchestrated by a language model that decides which commands to run. The project is actively maintained (v1.1.0) and invites contributions via GitHub and a Discord community.