AIDA Banner

AI-Driven Security Assessment

Autonomous Pentesting Agent

An agent that runs full security assessments end-to-end.
You define the scope. You review the findings.

Quick Start • What It Does • Installation • User Guide • Agent Tools • Discord

AIDA turns any LLM into an autonomous pentester capable of assessing web applications, APIs, and infrastructure. The agent reasons, understands application logic, executes commands in an isolated container, and documents every finding with the commands used.

AIDA Dashboard

Real Results

Claude + AIDA isn't just talk. It produces results that end up in CVE databases.

ID	Project	Description
CVE-2026-49869	kestra-io/kestra	Unauthenticated RCE via auth bypass + OS command injection (CWE-78/184/287/918)
CVE-2026-50189 / GHSA-xfvv-ggvq-pchh	appsmithorg/appsmith	RCE via newline injection in env variable endpoint
CVE-2026-32034	openclaw/openclaw	Insecure HTTP permits hijacking
CVE-2026-49979 / GHSA-vvxf-f8q9-86gh	appsmithorg/appsmith	SSRF via SMTP test endpoint — internal port scanning

More under responsible disclosure — awaiting publication.

What It Does

AIDA was built to give your AI everything a pentester needs to work.

A fully equipped execution environment. A Docker container loaded with Linux pentesting tools — nmap, sqlmap, ffuf, nuclei, and anything else it needs. If a tool is missing, the agent installs it.

Custom exploitation via Python. The agent generates and executes Python scripts on the fly — custom payloads, encoding tricks, protocol quirks, or any logic that off-the-shelf tools can't handle.

Burp-level HTTP control. The agent sends and manipulates HTTP requests directly — headers, cookies, body, auth tokens. Stored credentials are auto-injected via placeholders. Same power as Burp Repeater, without the UI overhead.

A persistent notebook. The agent logs what it knows about the application, maps attack paths, records observations, flags interesting behaviors, and documents every confirmed vulnerability — commands used, raw output, full context. Stop an engagement and resume it days later for retest, deeper analysis, or handoff.

Where is the real pentester?

You review. The AI hands you findings with full context notes, commands, reproduction steps, and the reasoning that led there. You reproduce, triage, prioritize, and report.

Your expertise stays where it matters. The grunt work runs on its own.

Quick Start

Prerequisites: Docker Desktop + any AI client (Claude, Gemini, GPT...)

git clone https://github.com/Vasco0x4/AIDA.git
cd AIDA
./start.sh

Dashboard: http://localhost:31337

./start.sh --dev — hot reload for contributors ./start.sh --lan — share across your local network (HTTPS, self-signed) ./start.sh --domain x.com — public deploy with Let's Encrypt

Launch the Agent

# Launches Claude Code by default
python3 aida.py --assessment "target-corp"

# Auto-detect Claude, Codex, Kimi, or Qwen
python3 aida.py --assessment "target-corp" --cli auto

# Force OpenAI Codex
python3 aida.py --assessment "target-corp" --cli codex

# No confirmation prompts
python3 aida.py --assessment "target-corp" --yes

Define Your Scope

Load assessment 'target-corp' and start the pentest on https://example.com
Scope: all subdomains, authentication flows, API endpoints
Exclude: brute-force on /login

Full setup for all AI clients → INSTALLATION.md

Supported Models

AIDA is model-agnostic. Any LLM with tool-calling support works.

Client	Setup
Claude Code	`python3 aida.py` (automatic)
OpenAI Codex CLI	`python3 aida.py --cli codex` (automatic)
Kimi Code CLI	`python3 aida.py` (automatic)
External API (OpenAI-compatible)	`python3 aida.py --base-url`
Claude Desktop	MCP config
ChatGPT Desktop	MCP config
Gemini CLI	MCP config

The smarter the model, the deeper the engagement. Swap models without changing anything else.

Agent Tools

Tool
`execute()`	Run any command in the pentesting container
`scan()`	nmap, gobuster, ffuf, nikto, dirb
`subdomain_enum()`	Subdomain discovery
`ssl_analysis()`	TLS/SSL audit
`tech_detection()`	Technology fingerprinting
`python_exec()`	Execute Python in the container
`http_request()`	HTTP client with credential auto-substitution
`add_card()`	Log a finding — CVSS 4.0 auto-scored
`credentials_add()`	Store credentials, auto-injected via `{{PLACEHOLDER}}`

Built-in aida-pentest container (~2 GB, starts automatically). Plug in Exegol for 400+ tools — switchable anytime from the dashboard.

Full reference → MCP_TOOLS.md

What's New in v1.1.0

Authentication — JWT, admin/user roles, first-run setup wizard
PDF reports — one-click export per assessment
CVSS 4.0 — automatic scoring on every finding
Attack timeline — auto-generated per engagement
Notifications — Telegram, Slack, Email with optional PDF attachment
Assessment templates — start from predefined methodologies
aida-pentest container — lightweight built-in environment, no Exegol required
python_exec + http_request — advanced execution tools
Cross-assessment findings view — aggregate and filter findings across all engagements
Security hardening — Docker socket proxy, path traversal prevention, localhost-only DB

!! Run locally or on your LAN. Do not expose the dashboard to the public. !!

Documentation


INSTALLATION.md	Full setup — all AI clients
USER_GUIDE.md	Platform usage guide
MCP_TOOLS.md	Agent tool reference
ARCHITECTURE.md	Technical deep dive

Contributing

AIDA is actively developed.

Planned:

OWASP testing guidelines integration
Multi-agent mode — specialized sub-agents per phase
Active Directory / internal network module
Enhanced CLI capabilities

Issues and PRs welcome → GitHub Issues

License

AGPL v3 — free and open source.

Credits

Anthropic MCP — the tool-calling protocol powering agent actions
Exegol — supported as alternative container
The security community for the open-source tooling

Questions or feedback? Join the community on Discord

⭐ Star on GitHub • 💬 Discord • Report Bug • Contact