In short
HexStrike AI is an open-source Model Context Protocol (MCP) server that lets AI agents run more than 150 security tools, such as nmap, nuclei and sqlmap, on their own. Security teams use it for penetration testing and bug bounty work. Attackers have discussed using it to find and exploit vulnerable systems faster.
Key takeaways
- What it is
- An open-source MCP framework, written in Python and published on GitHub under the MIT license, that gives AI assistants direct access to more than 150 security tools.
- How it works
- An AI client sends a task over MCP. HexStrike AI turns it into tool calls such as an nmap scan or a sqlmap test, runs them and returns the results so the model can pick the next step.
- Why it matters
- It shortens the time between a vulnerability disclosure and mass scanning. Check Point reported forum posts about using it against Citrix NetScaler flaws within hours of disclosure.
- What Odin Research saw
- 118 HexStrike AI endpoints in the reports covered here, none confirmed as hostile. Most sit on generic Python web servers on ports 5000, 8080 and 80, which also host many harmless apps.
- What to do first
- Patch internet-facing appliances within days, not weeks. Hunt for MCP tool calls that launch scanners from hosts that are not approved for testing.
What is HexStrike AI?
HexStrike AI is an open-source Model Context Protocol (MCP) server that lets AI assistants such as Claude, GPT and Copilot run more than 150 security tools on their own, from nmap port scans to sqlmap injection tests. Penetration testers use it to automate routine work, and attackers have discussed using it to exploit new flaws faster.
The project is published on GitHub as HexStrike AI MCP Agents, under the MIT license. The project describes itself as an MCP server for AI agents, with more than 150 tools and a dozen built-in agents for tasks such as reconnaissance, vulnerability scanning, bug bounty work and capture-the-flag challenges.
The key idea is the AI agent loop. A human gives the model a goal, for example "test this web application". The model then chooses tools, reads their output and decides what to run next. HexStrike AI supplies the tools and the plumbing between the model and the operating system.
That makes HexStrike AI a dual-use tool. It is built for penetration testing, and many installs belong to students, researchers and security teams. The same automation also helps attackers who want to scan and exploit many targets with little effort.
How HexStrike AI works
The Model Context Protocol is an open standard that lets an AI assistant call outside tools through a common interface. HexStrike AI uses it to hand security tools to the model.
HexStrike AI has three parts:
- The AI client. Any assistant that speaks the Model Context Protocol, such as Claude Desktop, Cursor or VS Code Copilot. The client holds the conversation and the model's reasoning.
- The MCP bridge. A small Python process (
hexstrike_mcp.pyin the project) that publishes each security tool as an MCP function, for example an nmap scan. The model sees these functions and can call them by name. - The API server. A Python web service (
hexstrike_server.py) that receives each call, runs the real tool on the host and returns the output. The project's examples use port 8888; installs seen on the internet often use 5000, 8080 or 80.
When the model calls a tool, the request travels as an MCP tools/call message to the bridge, then over HTTP to the API server. The result goes back the same way. Because the model reads every result, it can chain steps: discover open ports, fingerprint the web server, run a template scan, then try an injection test on the parameter that looked weak.
To learn how MCP itself should be locked down inside your own organization, see Securing MCP Servers and Tool Access.
HexStrike AI requirements, installation and popularity
According to the project's README, HexStrike AI needs Python 3.8 or later and is installed from its GitHub repository with pip and a requirements file. Most of the security tools it calls must be installed separately, which is why many users run it on a Kali Linux or similar testing machine. The user starts the API server, then points an MCP client such as Claude Desktop, Cursor or VS Code Copilot at the bridge through the client's MCP configuration. The README's examples use port 8888 for the API server.
The README credits the author 0x4m4 and labels the current edition v6.0. The repository had about 13,800 stars and 2,800 forks on GitHub on 21 November 2026, which makes it one of the most visible AI security tool projects. That popularity is one reason it appears in scans of internet-facing servers.
Is it legal to use? The software is MIT-licensed and legal to download. Running it against systems you do not own or have written permission to test is not. Keep its use inside an authorized, documented test scope.
How attackers use HexStrike AI
In September 2025, Check Point Research reported that threat actors on underground forums were discussing HexStrike AI within hours of a Citrix disclosure. The posts talked about using it to scan for and exploit three NetScaler ADC and Gateway flaws published on 26 August 2025:
- CVE-2025-7775, a memory overflow that can lead to remote code execution (CVSS 4.0 score 9.2, listed in CISA's Known Exploited Vulnerabilities catalog);
- CVE-2025-7776, a memory overflow that can cause erroneous behavior and denial of service (CVSS 4.0 score 8.8);
- CVE-2025-8424, improper access control on the NetScaler management interface (CVSS 4.0 score 8.7).
According to Check Point, some posters claimed the framework cut the time to exploitation from days to under ten minutes, and some offered lists of vulnerable hosts. These are claims made in forum posts, not measured results. They still show the direction of travel: AI orchestration lowers the skill and time an attacker needs between a patch release and mass exploitation.
For the response side, see How to respond to actively exploited vulnerabilities and our page on Exploit Public-Facing Application (T1190).
What Odin Research observed
Odin Research has tracked HexStrike AI servers alongside other AI offense frameworks in several studies. The numbers below come from those reports; addresses stay withheld because none of these hosts is confirmed as hostile.
- 118 endpoints, none confirmed. In Monograph 11, 99 HexStrike AI endpoints were suspected and 19 were candidates. Many installs show only a generic Python web page, so nothing on the page proves the framework is there.
- Generic ports, crowded neighbours. Monograph 04 counted 56 records on 53 IP addresses, most on ports 5000 and 8080, on Shenzhen Tencent, DigitalOcean and cognetcloud networks.
- Lookalikes are common. One suspected host turned out to be a harmless copy of a security training portal that only shared a Python web server header with HexStrike AI. Another showed a Chinese login page titled "Smart Testing Tool".
- Shared infrastructure. HexStrike AI often runs next to other offensive tools. Monograph 11 found it on the same server as DeepAudit, Monograph 17 saw a HexStrike AI tool bridge in the same cloud subnet as a DNSLog collector, a VShell listener and a CyberStrikeAI engine, and Monograph 19 found it beside Mythic and Havoc C2 servers.
The lesson for defenders: a HexStrike AI banner alone is weak evidence. Treat these servers as hunting leads and look for a second, independent sign before blocking.
How to detect HexStrike AI
Detection works best when it combines several signals:
- Process names on your own hosts. Alert when
hexstrike_server.pyorhexstrike_mcp.pyruns outside approved testing machines. - MCP tool calls that launch scanners. A
tools/callrequest that names nmap, nuclei, sqlmap, ffuf or gobuster, sent from a workstation or server that has no testing role, deserves a look. - Bursts that no human would type. Agent-driven scans issue many requests in short, regular bursts and try one injection pattern after another against the same parameter.
- Not the port alone. Ports 5000 and 8080 with a Python web server header are far too common to block on.
Odin Research publishes ready-to-use rules for these patterns. Monograph 11 includes a Suricata signature for HexStrike AI MCP tool-call traffic and a YARA rule for HexStrike AI server files, and Monograph 17 adds a Sigma rule for offense daemons started outside approved red-team hosts. All rules are on the report pages and in the Defense Kit.
How to defend against AI-assisted exploitation
- Patch exposed edge devices first. Prioritize flaws in the KEV catalog and appliances on the internet. AI-driven scanning narrows the window you have.
- Know your attack surface. You cannot patch what you have not found. External Attack Surface Management shows which services an automated scanner would see.
- Rate-limit and watch for agent patterns. Web application firewalls and API gateways can slow automated probing and flag rapid, systematic parameter testing.
- Govern your own AI tools. If your teams use HexStrike AI or similar frameworks, keep them on dedicated testing hosts, log every tool call and require written scope. See the AI Agent Governance Policy Template.
HexStrike AI compared with other AI pentest frameworks
| Framework | How it drives tools | Typical ports seen | Notes |
|---|---|---|---|
| HexStrike AI | MCP server exposing 150+ command-line tools to any MCP client | 5000, 8080, 80, 8888 | Generic Python web pages make installs hard to confirm |
| CyberStrikeAI | Own platform with MCP tools and an agent listener | 5001 to 5003 and 8082 on one host | The multi-port layout is a stronger signal than any banner |
| PentAGI | Autonomous agents that start Kali containers | Web console ports | Dual-use; a login page alone proves little |
| ARTEX | Multi-agent system with a task API | 8787 and forks on other ports | Linked by vendors to real intrusions; see our ARTEX profile |
Frequently asked questions about HexStrike AI
What is HexStrike AI?
HexStrike AI is an open-source Model Context Protocol (MCP) server that lets AI assistants such as Claude, GPT and Copilot run more than 150 security tools on their own, for penetration testing, bug bounty and research work.
Is HexStrike AI malware?
No. HexStrike AI is a legitimate, MIT-licensed security tool built for authorized testing. It is dual-use: attackers can run the same automation against systems they have no permission to test, which is why defenders track it.
Is HexStrike AI legal to use?
Yes, when you use it on systems you own or are authorized to test. The software is MIT-licensed and free to download, but scanning or exploiting systems without written permission is illegal in most countries.
How do I install HexStrike AI?
The project's README says it needs Python 3.8 or later. You install it from the GitHub repository with pip and its requirements file, install the security tools it calls, start the API server and add the MCP bridge to an MCP client such as Claude Desktop or Cursor.
Which tools does HexStrike AI use?
The project says it integrates more than 150 tools. Common examples are nmap for port scanning, nuclei for template scans, sqlmap for SQL injection tests, and ffuf and gobuster for content discovery.
Which AI models work with HexStrike AI?
Any client that supports MCP. The project documents Claude Desktop, Cursor, VS Code Copilot and Roo Code, and the model behind the client can be Claude, GPT or another large language model.
How have attackers used HexStrike AI?
Check Point Research reported in September 2025 that forum users discussed using HexStrike AI to scan for and exploit Citrix NetScaler flaws, including CVE-2025-7775, within hours of their disclosure. The speed claims came from the posters themselves.
How can I detect HexStrike AI in my network?
Look for hexstrike_server.py or hexstrike_mcp.py processes outside approved testing hosts, for MCP tool calls that launch scanners, and for bursts of systematic probing. Odin Research publishes Suricata, YARA and Sigma rules for these patterns in its Defense Kit.
Should I block every HexStrike AI server I find?
No. Most installs Odin Research has seen run on generic Python web servers that many harmless apps also use. Confirm a second, independent signal of hostile use before blocking an address.
Sources
- 0x4m4, HexStrike AI MCP Agents repository, GitHub. github.com/0x4m4/hexstrike-ai (checked 21 November 2026)
- Check Point, "Hexstrike-AI: When LLMs Meet Zero-Day Exploitation", 2 September 2025. blog.checkpoint.com/executive-insights/hexstrike-ai-when-llms-meet-zero-day-exploitation/ (checked 21 November 2026)
- NVD, CVE-2025-7775. nvd.nist.gov/vuln/detail/CVE-2025-7775 (checked 21 November 2026)
- NVD, CVE-2025-7776. nvd.nist.gov/vuln/detail/CVE-2025-7776 (checked 21 November 2026)
- NVD, CVE-2025-8424. nvd.nist.gov/vuln/detail/CVE-2025-8424 (checked 21 November 2026)
- Model Context Protocol specification. modelcontextprotocol.io/ (checked 21 November 2026)
Related tools and terms
What is CyberStrikeAI?
CyberStrikeAI is a public AI-driven offensive security platform. It spreads its operator interface, API and agent listener across several ports on one host and exposes tools through MCP.
AI offensive toolWhat is PentAGI?
PentAGI is an open-source autonomous penetration testing system that coordinates AI agents and security tools in containers.
AI security toolWhat is DeepAudit?
DeepAudit is an open-source AI code auditing tool. Application security teams use it to find flaws in source code.
AI security toolWhat is AutoCVE?
AutoCVE is an AI-assisted tool that turns new vulnerability advisories into HTTP checks.
AI security toolWhat is Cairn?
Cairn is a reconnaissance tool that keeps an asset graph built from DNS zones, TLS certificates and HTTP banners.
AI security toolWhat is AutoHunter?
AutoHunter automates headless bug-bounty reconnaissance of subdomains, API routes and query parameters.
Odin Research
Get the rules, not just the definition
Every Odin Research report ships with verified indicators and detection rules you can load today.