PPhying中文
Curated security research, vulnerabilities, advisories and tools for practitioners.

StrikeAgent AtkBrain-Flash: Autonomous AI Penetration Testing & Red Teaming Platform

github.com | tool | #ai-security | #pentesting | #claude-code | #automation | #red-team | #ai-agents | #kali-linux
🔗 Original ↗

Summary

StrikeAgent AtkBrain-Flash: autonomous AI pentesting platform pairing Claude Code with Kali for attack-graph-driven exploitation, dual-agent orchestration, lateral-movement mapping and reporting.

Why it matters

Demonstrates practical autonomous agent execution in offensive cybersecurity, bridging high-level LLM reasoning with real-world penetration testing toolchains for automated initial access.
Published
Collected

📦 Full-Text Content Snapshot

R2 Archive · Full Article Preserved187 lines · 12.7 KB
📄 Raw .md⬇️ Download .md
Triad: Self-looping · Self-monitoring · Self-evolving
Triad: Self-looping · Self-monitoring · Self-evolving

This project was developed by the Yean-Sec team to explore autonomous AI penetration testing capabilities, aiming to build a solution with authentic engineering philosophies rather than generic, templated AI wrappers.

While the overall platform may appear straightforward on the surface, it has undergone multiple rounds of rigorous real-world engagements and embodies extensive engineering effort. Crucial design details abound: solving the dilemma of "expanding the attack surface without losing focus or drifting off-target" alone required substantial refinement (in the Flash edition). Furthermore, to prevent AI-generated vulnerabilities from suffering from hallucination or exaggeration, the team engineered red-team secondary verification and secondary rating mechanisms, ensuring findings are what-you-see-is-what-you-get and directly actionable.

The team's internal Pro version has already been deployed across dozens of engagements and over a thousand authorized external targets, focusing entirely on external initial access.

Architecture

The console dispatches targets; the attack graph drives the autonomous execution loop. The agent executes a complete attack round before prompting the operator, asking for guidance only when encountering a blocker or reaching the cycle threshold; manual commands entered into the dialogue immediately interrupt the active round. Upon concluding an engagement, operational tactics are distilled back into the memory base. The runtime is powered by Pi, while the LLM provider is dynamically configured via settings JSON. Agent tests connect directly to real IPs by default; traffic routes through proxy pools only when explicitly configured and enabled.

StrikeAgent_AtkBrain-Flash Architecture
StrikeAgent_AtkBrain-Flash Architecture

UI Showcase

Platform Overview:

Platform Overview
Platform Overview

Home & Project Creation: Dual entry points for Agent Testing and CTF challenges. Targets can be input individually or imported in batches via .txt / .csv alongside contextual parameters. CTF challenges support Web, PWN, Reverse Engineering, Misc, and Cryptography categories with attachment uploads.

Project Creation
Project Creation

Project Console: Real-time attack graph, execution timeline, verified vulnerability feed, and interactive agent dialogue.

Project Console
Project Console

Vulnerability Base:

Vulnerability Base
Vulnerability Base

Delivery Reports

Report Template Sample:

Open Interactive HTML Report

Report Cover
Report Cover
Executive Summary
Executive Summary
Critical Attack Paths
Critical Attack Paths
Asset Profile
Asset Profile

Benchmark Leaderboard

Ranked #1 on Tsecbench v1 (StrikeAgent_AtkBrain-Flash, score: 97.89 / 100), the intelligent offensive-defensive security agent evaluation led by Tencent Security's Yunding Lab.

Tsecbench Leaderboard
Tsecbench Leaderboard

Security Hardening Efforts

To prevent unauthorized exposure, default installations conceal login pages, API endpoints, and plaintext secrets:

  • Randomized Access Entry: An 8-character random path is written strictly to local backend/data/security_entry (0600 permissions); it is never exposed over the web or APIs. Run python -m atkbrain.panel on the host to print the URL.
  • No Entry, No Subdirectory: Root :2334/ renders only a static introduction page; /login, /api, /.env, etc., unconditionally return HTTP 404.
  • Anti-Interception / Anti-Replay / Anti-Brute-Force: Browser encrypts authentication payloads via RSA-OAEP before POSTing; one-time tickets (120s TTL); rate limiting and IP/username lockouts.
  • Credential Storage: Passwords hashed via Argon2id; local admin_bootstrap.secret (0600) accessible only via host shell.
  • Session Protection: HttpOnly cookies; no query tokens (?token=); Swagger documentation disabled; HTTPS on :2334 by default.

Environment & Installation

Choose between Docker or a native Linux systemd setup. Docker packages the console, Pi, and standard recon utilities into containers. The native Linux setup runs via systemd without virtualization. Do not run both simultaneously on ports 2333 / 2334.

LLM models are not configured during installation. Once the console is running, configure Pi's model JSON via settings. .env stores only local port bindings and administrative secrets; do not store vendor API keys in .env. Persistent data lives under backend/data/ (gitignored), so upgrades or service reinstalls will not overwrite databases, entry paths, or saved model settings.

Docker Deployment

Requires Docker and Docker Compose. Execute from the repository root (containing compose.yaml and .env.example). Linux uses host networking; remote clients must be able to reach host port 2334. Docker Desktop on macOS lacks host networking, so append deploy/compose.mac.yaml to bind the console to 127.0.0.1.

bash
git clone https://github.com/Yean-Sec/StrikeAgent_AtkBrain-Flash
cd StrikeAgent_AtkBrain-Flash
cp .env.example .env
# Default .env values suffice. Initial login: admin / admin (change required upon login).
# Do NOT enter model API keys here.

# Linux
docker compose -f compose.yaml -f deploy/compose.build.yaml up -d --build --wait

# macOS
# docker compose -f compose.yaml -f deploy/compose.build.yaml -f deploy/compose.mac.yaml up -d --build --wait

--build builds image strikeagent-atkbrain-flash:console from Dockerfile.console, while --wait holds until healthchecks pass. Containers created: strikeagent-atkbrain-flash (API) and strikeagent-atkbrain-caddy (HTTPS).

bash
docker compose -f compose.yaml ps
docker compose -f compose.yaml exec atkbrain python -m atkbrain.panel
# Prints the random https entry URL and active password (visible only on this host)

# Upgrade: pull latest code in repository root and rebuild
docker compose -f compose.yaml -f deploy/compose.build.yaml up -d --build --wait

# Stop services
docker compose -f compose.yaml -f deploy/compose.build.yaml down

Native Linux Deployment

Runs directly on Linux without containers. Requirements: root privileges, systemd, Python 3.11+ (/usr/bin/python3), Node.js 20+, npm, OpenSSL. System units bind strictly to system python3; dependencies must be installed into this interpreter rather than an unlinked venv.

Debian / Ubuntu:

bash
sudo apt-get update
sudo apt-get install -y python3 python3-pip python3-venv nodejs npm openssl ca-certificates   gcc pkg-config libffi-dev   libcairo2 libpango-1.0-0 libpangocairo-1.0-0 libgdk-pixbuf-2.0-0

RHEL / TencentOS:

bash
sudo dnf install -y python3 python3-pip nodejs npm openssl ca-certificates gcc pkgconf-pkg-config libffi-devel cairo pango gdk-pixbuf2

Ensure Node.js is version 20+. Then install Python dependencies, frontend, and Pi:

bash
git clone https://github.com/Yean-Sec/StrikeAgent_AtkBrain-Flash
cd StrikeAgent_AtkBrain-Flash
cp .env.example .env

python3 -m pip install -r backend/requirements.txt
npm install -g @earendil-works/pi-coding-agent
cd frontend && npm install && cd ..

sudo scripts/atkbrain-up.sh

atkbrain-up.sh provisions and starts two systemd units:

  • atkbrain-flash-backend.service: API service on 2333
  • atkbrain-flash-frontend.service: Self-signed HTTPS on 2334, reverse proxying to 2333
bash
scripts/atkbrain-panel.sh
# Equivalent to: cd backend && python3 -m atkbrain.panel

sudo scripts/atkbrain-backend.sh status
sudo scripts/atkbrain-frontend.sh status
sudo scripts/atkbrain-backend.sh logs
sudo scripts/atkbrain-frontend.sh logs

Logs are recorded in backend/data/logs/backend.log and frontend.log.

Upgrading: Pull new commits, reinstall dependencies, and restart services:

bash
python3 -m pip install -r backend/requirements.txt
cd frontend && npm install && npm run build && cd ..
sudo scripts/atkbrain-backend.sh restart
sudo scripts/atkbrain-frontend.sh restart

To stop: sudo scripts/atkbrain-backend.sh stop and sudo scripts/atkbrain-frontend.sh stop. To remove units while preserving data, execute uninstall on both scripts.

Note on tools: The Docker image includes standard network reconnaissance tools. Native Linux instances provide only the console and Pi runtime; utilities like nmap must be installed separately on the host and available in root's PATH.

Post-Deployment Workflow

Open the HTTPS URL printed by panel in your browser (accept the self-signed certificate). The console is on :2334, and API on :2333. Bare :2334/ displays the overview page; append the generated random path to access the login prompt.

Login with initial credentials admin / admin, and update your password immediately when prompted.

Configure LLM Models: Navigate to Settings → Models, select OpenAI or Anthropic, and configure the model name, Base URL, and API Key. OpenAI uses Chat Completions; Anthropic uses Messages API. Settings are saved to backend/data/pi-models.json. Click "Test Connectivity" to send a minimal verification request.

Password / Entry Recovery: Rerun panel. If plaintext backups are missing, configure ATKBRAIN_ADMIN_PASSWORD and ATKBRAIN_ADMIN_PASSWORD_RESET=1 in .env, run docker compose ... up -d or sudo scripts/atkbrain-backend.sh restart, verify access, and revert RESET to false.

License & Disclaimer

AGPL-3.0-only — Free for personal and open-source use. Commercial licensing inquiries: gavenmiya@outlook.com. Unauthorized commercial distribution is strictly prohibited. See LICENSE for full legal terms.

Authorized Engagements Only. Use of this software is strictly restricted to explicitly authorized targets. Users assume all responsibility for their operations. The platform hardcodes testing blocks against critical domains (e.g. .gov, .edu) and enforces strict guardrail filters. The platform does not bundle bypasses, proxy relays, or evasion payloads!

📦 Full-Text Markdown Snapshot archived by Phying News · Original source: github.com

Related coverage

tool · github.com
SpecterOps Skills: Open-Source LLM Agent Skills & MCP Plugin Marketplace for Red Teaming and AppSec
As AI-assisted offensive security evolves from simple prompting to multi-agent architectures, SpecterOps standardizes security craft into declarative skills and MCP interfaces, allowing operators to safely orchestrate complex red team tradecraft and reverse engineering directly through LLM agents.
blog · aikido.dev
Aikido launches external network pentesting
TL;DR: Aikido now pentests your external network. Attackers already test your perimeter every day, and with AI agents doing the grunt work, they're faster than ever. Now you can run the same test first. Give our agents one domain and they'll map the internet-facing systems you run and attack them from the outside, exactly like a malicious actor would.
blog · aikido.dev
Adding Aikido AI pentesting to your existing security stack
How to add Aikido's AI pentesting on top of the SAST, SCA and cloud tools you already run, from your first pentest through whitebox testing, continuous testing and remediation.
blog · certik.com
CertiK Skills: Bringing Blockchain Security Intelligence Into AI Agents
CertiK open-sourced four AI agent skills that plug its blockchain security intelligence into Claude Code, Codex, and Cursor, letting agents query wallet risk, AML screening, and EVM transaction forensics natively.
blog · hacktron.ai
How can we make AI hack like a human?
Explores how elite human hackers reason through complex systems and lays out design principles for building AI agents that can emulate that investigative process to find real bugs.
tool · github.com
Brotli Client Checker: Test Server for Brotli Decoder Vulnerability (CVE-2026-111497)
Provides security teams and developers with a simple, zero-dependency mechanism to test whether backend microservices, crawlers, and HTTP clients are susceptible to the widespread Node2Shell/HTTP2Shell Brotli decoder flaw.
Back