Skip to content
NewCybercrime Report 2026 · Eight Chapters · A Post Is Not An Incident → Explore The ChaptersCybercrime Report 2026 → Explore

Odin Research / What is / AI tools and agents

What Is CyberStrikeAI? The AI Pentest Platform and How to Detect It

An AI-native offensive security platform that pairs language models with more than 100 security tools: how it works, what researchers have tied it to, and how to spot it.

By
Odin by Brandefense
Updated
21 Nov 2026
Reading time
7 min read
Type
AI offensive security platform

In short

CyberStrikeAI is an open-source, AI-native security testing platform written in Go. It connects large language models to more than 100 security tools through agents and the Model Context Protocol (MCP). The project says it is for authorized penetration testing. Researchers have linked servers running it to scanning of FortiGate devices.

Key takeaways

What it is
An open-source AI pentest platform on GitHub, written in Go under the Apache 2.0 license. It wraps more than 100 security tools as agent tools.
How it works
A web console sends a goal to AI agents. The agents pick tools through MCP, read the output and plan the next step, using any OpenAI-compatible model.
Why it matters
Team Cymru reported in March 2026 that a server Amazon linked to FortiGate intrusions ran CyberStrikeAI, and counted 21 IP addresses running it in early 2026.
What Odin Research saw
1,290 CyberStrikeAI records on 528 IP addresses in Monograph 04, 73.4% on one hosting provider, many with a distinct four-port layout.
What to do first
Keep management interfaces off the internet, enforce MFA on edge devices and hunt for hosts that reach ports 5001, 5002, 5003 and 8082 on one address.

What is CyberStrikeAI?

CyberStrikeAI is an open-source, AI-native security testing platform written in Go. It lets large language models drive more than 100 security tools through agents and the Model Context Protocol (MCP), from scanning to post-exploitation. The project says it is for authorized testing, but researchers have tied it to real intrusion activity.

The GitHub README calls it a platform for AI-native cybersecurity. It combines AI agents, MCP-based tools, a knowledge base, workflows and attack-chain analysis in one web console. Documentation is in English and Chinese.

CyberStrikeAI belongs to a new class of AI offense frameworks, next to HexStrike AI, PentAGI and ARTEX. A human sets a goal. The model then chooses tools, reads results and decides what to run next. That makes it a dual-use tool: useful for penetration testing, and just as useful to an attacker who wants speed and scale.

How CyberStrikeAI works

The Model Context Protocol is an open standard for connecting AI applications to outside tools and data. According to the README, CyberStrikeAI has five main layers:

  1. Tool recipes. More than 100 curated YAML recipes wrap command-line tools across the kill chain, from network scanners to forensics and post-exploitation.
  2. MCP. Tools are exposed over MCP through HTTP, stdio and SSE transports, with support for external MCP servers. Monograph 11 notes that its MCP tools/call schemas include automated sqlmap runs and upload probes.
  3. Agents and roles. It supports a single agent and three multi-agent modes: Deep, Plan-Execute and Supervisor. More than 12 predefined testing roles limit which tools each agent may use.
  4. Knowledge and skills. A vector knowledge base with reranking and on-demand agent skills feed the model context.
  5. Operations modules. Graph workflows, asset and vulnerability management, WebShell and C2 modules, and chatbot integrations.

Models connect through an OpenAI-compatible setting. The README's examples include OpenAI and DeepSeek. The server listens on port 8080 by default and the launch script serves HTTPS unless told otherwise. The README lists Go 1.25 or later and Python 3.10 or later as requirements.

To learn how MCP tool access should be locked down in your own environment, see Securing MCP Servers and Tool Access.

CyberStrikeAI origin and popularity

Team Cymru reported that the repository's first commit was on 8 November 2025. On 21 November 2026, the repository sat under the AIPentest organization on GitHub, with about 8,900 stars and 2,000 forks. Its quick-start clone command points to the account Ed1s0nZ.

Team Cymru looked at the developer behind that account. It reported that on 19 December 2025 the developer posted CyberStrikeAI to the Starlink Project of Knownsec 404. It also reported that on 5 January 2026 the developer added a CNNVD contribution award to the GitHub profile and later removed it. Based on this, Team Cymru assessed that the developer has some ties to the Chinese government. That is Team Cymru's assessment, not a confirmed attribution of any intrusion.

Is it legal to use? The software is free to download. The README says to use it only on systems you own or are explicitly authorized to test, and flags its WebShell and C2 features as high risk.

How attackers use CyberStrikeAI

Team Cymru published its research on 2 March 2026. It said it first learned of CyberStrikeAI from an Amazon threat intelligence report about an AI-augmented actor that accessed FortiGate devices at scale. Team Cymru found that an IP address shared by Amazon ran a CyberStrikeAI banner. Its NetFlow data showed that address talking to targets, including FortiGate devices.

Team Cymru then counted 21 unique IP addresses running CyberStrikeAI between 20 January and 26 February 2026. Most sat in China, Singapore and Hong Kong. It described this as a sharp rise in operational use after the repository appeared. Team Cymru said the tool is likely to be used by Chinese state-sponsored groups, but it did not confirm use by any named group.

For defenders, the pattern matters more than the brand. AI orchestration cuts the time and skill needed to scan many edge devices and try weak or reused credentials. See How to respond to actively exploited vulnerabilities.

What Odin Research observed

CyberStrikeAI is one of the largest AI offense frameworks in Odin Research's studies. Addresses stay withheld unless a report published them.

  • A dominant, concentrated fleet. Monograph 04 counted 1,290 CyberStrikeAI records on 528 IP addresses. Together with ARTEX, it made up 95.6% of 1,874 AI offense endpoints. PEG TECH INC carried 947 of its records (73.4%).
  • A four-port fingerprint. Ports 5001, 5002, 5003 and 8082 each had 237 listeners, and 236 hosts bound all four at once. Monograph 02 treats this layout, not any single banner, as the useful signal. It also found 99 records with a default Ubiquiti certificate name that never showed the four-port layout, so that name is only a lead.
  • A co-hosted workbench. Monograph 04 found an open directory of 363 files on a server running CyberStrikeAI v1.7.24 next to ARTEX. The files showed both tools wired to a DeepSeek model, search-engine keys and Telegram alerts. A short script was written to restore the admin password hash and last-updated time in the local database, a form of indicator removal.
  • Growth across studies. Monograph 11 counted 1,720 endpoints on 1,142 hosts, with the highest confirmed share of any framework at 75.0%.
  • Busy neighbors. Monograph 17 saw a CyberStrikeAI engine in the same cloud subnet as a DNSLog collector, a VShell listener and a HexStrike AI bridge, on separate, unlinked addresses. Monograph 19 found it on hosts beside DarkSword.

A running instance shows the software is there. It does not show who runs it or whether they are authorized.

How to detect CyberStrikeAI

  • The four-port layout. A device that reaches 5001, 5002, 5003 and 8082 on the same address is a strong lead, even at low volume.
  • Certificates. Some nodes present CN=CyberStrikeAI. Treat a default certificate as a hunting lead, not a block trigger.
  • Exposed workspaces. HTTP requests for /CyberStrikeAI/ paths point to an exposed workspace or config.
  • Processes. On your own Linux hosts, alert on cyberstrikeai processes outside approved red-team machines.

Odin Research publishes rules for these signals. Monograph 04 has Suricata signatures for CyberStrikeAI workspace and token-store requests and a YARA rule for its configs and rollback script. Monograph 02 adds a Sentinel KQL hunt and a Sigma rule for the four-port layout, and Monograph 17 a Sigma rule for offense daemons. All are in the Defense Kit.

How to defend against AI-driven attacks like CyberStrikeAI

  1. Close the front door. Remove edge-device management interfaces from the internet and require MFA. Team Cymru's case started with FortiGate devices.
  2. Patch fast. AI-driven scanning narrows the window between disclosure and mass exploitation. Prioritize flaws in the KEV catalog.
  3. Rate-limit and watch for agent bursts. Systematic, rapid probing of one parameter after another is a sign of automation.
  4. Govern your own use. If your team runs CyberStrikeAI, keep it on dedicated hosts, log every tool call and keep its console private.

CyberStrikeAI compared with other AI offense frameworks

Counts come from Monograph 11.

FrameworkDesignEndpointsTypical ports
CyberStrikeAIGo and Python platform with role-based agents wrapping 100+ tools1,7205001 to 5003, 8080, 8082, 443
ARTEXPython multi-worker swarm with a task queue2,1218787, 443, 80, 7878
PentAGIGo controller that starts Docker-in-Docker Kali workers3408443, 8666, 443
HexStrike AIPython server that exposes nmap, nuclei, sqlmap and ffuf as MCP tools1185000, 80, 8080, 8000
DeepAuditLLM workbench for auditing source code618000, 3000, 80

Frequently asked questions about CyberStrikeAI

What is CyberStrikeAI?

CyberStrikeAI is an open-source, AI-native security testing platform written in Go. It connects large language models to more than 100 security tools through agents and MCP, for penetration testing and research.

Is CyberStrikeAI malware?

No. It is an Apache-2.0-licensed security tool that says it is for authorized testing. It is dual-use, and Team Cymru linked one server running it to activity against FortiGate devices.

Who made CyberStrikeAI?

The repository sits under the AIPentest organization on GitHub, and its clone instructions name the account Ed1s0nZ. Team Cymru assessed that this developer has some ties to the Chinese government.

Is CyberStrikeAI legal to use?

Downloading it is legal. Using it is legal only on systems you own or have written permission to test. The README itself limits use to authorized systems.

Which AI models does CyberStrikeAI support?

It connects to models through an OpenAI-compatible provider setting. The README's examples include OpenAI and DeepSeek. Odin Research saw one install wired to a DeepSeek model.

What ports does CyberStrikeAI use?

The README gives port 8080 as the default. Odin Research found many hosts that bind ports 5001, 5002, 5003 and 8082 together, and that four-port layout is a useful hunting signal.

Was CyberStrikeAI used in the FortiGate attacks?

Team Cymru reported that an IP address Amazon shared from its FortiGate investigation ran a CyberStrikeAI banner, and that NetFlow showed it talking to FortiGate devices. Team Cymru did not tie the tool to a named group.

How do I detect CyberStrikeAI?

Hunt for devices that reach ports 5001, 5002, 5003 and 8082 on one address, CN=CyberStrikeAI certificates, requests for /CyberStrikeAI/ paths and cyberstrikeai processes. Odin Research publishes Suricata, Sigma, YARA and KQL rules in its Defense Kit.

CyberStrikeAI vs HexStrike AI: what is the difference?

HexStrike AI is mainly an MCP server that hands security tools to an outside AI client. CyberStrikeAI is a full platform with its own console, agents, roles, knowledge base and C2 modules.

Sources

  1. CyberStrikeAI repository and README, GitHub. github.com/Ed1s0nZ/CyberStrikeAI (checked 21 November 2026)
  2. Team Cymru, "Tracking CyberStrikeAI Usage", 2 March 2026. www.team-cymru.com/post/tracking-cyberstrikeai-usage (checked 21 November 2026)
  3. Model Context Protocol specification. modelcontextprotocol.io/ (checked 21 November 2026)

Related tools and terms

All Odin Research explanations · Brandefense glossary

Odin Research

Get the rules, not just the definition

Every Odin Research report ships with verified indicators and detection rules you can load today.

Open the Defense Kit