In short
PentAGI is an open-source, self-hosted multi-agent system that runs penetration tests on its own. A user describes a target and goal in plain language. AI agents then plan the work, run tools such as nmap, Metasploit and sqlmap inside isolated Docker containers, and write a report. It is built for authorized testing.
Key takeaways
- What it is
- An MIT-licensed project from the vxcontrol GitHub organization: a team of specialized AI agents that plans and carries out penetration tests with 20+ built-in security tools.
- How it works
- A web console and GraphQL or REST API receive a goal. Agents such as Pentester, Coder and Searcher run tools in sandboxed Docker containers and store what they learn in PostgreSQL with pgvector.
- Why it matters
- It lowers the skill needed to run a multi-step attack. The same automation that helps testers can help an intruder scan and probe many targets.
- What Odin Research saw
- 340 PentAGI endpoints in Monograph 11, only 22 confirmed. One confirmed controller shared a server with an ARTEX port and a staged upload-bypass dropzone.
- What to do first
- Keep your own installs on localhost or behind a VPN, change the default admin password, and alert on container runtimes that launch scanners from non-testing hosts.
What is PentAGI?
PentAGI is an open-source, self-hosted multi-agent system for autonomous penetration testing. You describe a target and a goal in plain language. Specialized AI agents then plan the work, run security tools inside an isolated Docker sandbox and produce a report. The project says it is meant for authorized testing only.
The code lives in the vxcontrol organization on GitHub under the MIT license, with an end-user license agreement that covers acceptable use. The repository describes PentAGI as a fully autonomous AI agents system for complex penetration testing tasks.
PentAGI belongs to a new class of AI agent tools for offensive security. A model does not only answer questions. It picks the next tool, reads the output and decides what to try after that. That makes PentAGI useful for penetration testing teams and worth watching for defenders.
How PentAGI works
According to the project README, PentAGI combines several parts:
- A web console and APIs. Users work through a web interface or through REST and GraphQL APIs, both authenticated with Bearer tokens.
- A team of agents. The README names roles such as Primary Agent, Pentester, Coder, Searcher, Adviser and Reporter. Each handles one part of the job, from research to exploitation to writing up findings.
- A tool sandbox. All operations run in isolated Docker containers. The built-in suite has more than 20 security tools, including nmap, Metasploit and sqlmap.
- Long-term memory. Research results and approaches that worked are stored in PostgreSQL with the pgvector extension. An optional knowledge graph, built on Graphiti and Neo4j, tracks links between hosts, services and findings.
- Monitoring. Langfuse records model activity. Grafana, OpenTelemetry, VictoriaMetrics, Jaeger and Loki cover system health.
PentAGI works with more than ten cloud model providers, including OpenAI, Anthropic, Gemini, AWS Bedrock and Mistral. It also supports any OpenAI-compatible endpoint and local models through Ollama or vLLM. Reports come out as a web view, Markdown or PDF.
Odin Research's own analysis in Monograph 11 describes the deployed controller as a Go service with a GraphQL API. It starts Kali Linux worker containers through Docker-in-Docker, and an LLM decides which tool to call next.
PentAGI requirements, installation and popularity
The README lists Docker and Docker Compose, or Podman, as the base. The minimum system is 2 vCPU, 4 GB of RAM and 20 GB of free disk, plus internet access for images and updates.
The web interface listens on https://localhost:8443 with a self-signed certificate. The README says it binds to localhost by default. A fresh install creates a default local administrator account, and the README tells users to change its password at first login.
The repository had about 34,700 stars and 5,500 forks on 21 November 2026. That puts PentAGI among the most visible open-source AI pentest projects, and it explains why its consoles turn up in internet scans.
Is PentAGI legal? Yes, as software. Running it against systems you do not own or lack written permission to test is not.
Why PentAGI matters to defenders
PentAGI is not malware. The concern is what happens when an autonomous testing system runs without scope or approval.
- Speed and scale. Agents can chain discovery, fingerprinting and exploitation without a human typing each step.
- Exposed consoles. When an operator moves the console off localhost, a login page and an API appear on the internet. Each one is a potential entry point to a box full of offensive tools and model keys.
- Shared hosts. Odin Research often finds PentAGI on the same servers as other offense frameworks, which makes the whole host more interesting than any single banner.
The container step matters for detection. Starting tools inside containers maps to Container Administration Command (T1609) in MITRE ATT&CK.
What Odin Research observed
Odin Research has tracked PentAGI alongside other AI offense frameworks in five studies. Addresses are withheld unless a report confirms hostile use.
- Many panels, few confirmations. Monograph 11 counted 340 PentAGI endpoints on 333 hosts. Only 22 were confirmed, 4 were suspected and 314 were candidates. A responsive panel is not a confirmed deployment.
- A common certificate. The subject
CN=localhostappeared on 61 of the 340 endpoints. Upstream default installs presentCN=pentagiinstead. Monograph 02 also found PentAGI among the tools sharing genericCN=localhostcertificates, next to Sliver and AdaptixC2. - A shared offense host. One confirmed PentAGI controller in the United States sat beside an ARTEX port and an open directory. That directory held an Apache
.htaccessfile written to make uploaded images run as PHP, plus a GraphQL schema dump and a list of MCP tools. These files were staged, and nothing shows they worked against a real target. - Hosting and ports. Monograph 04 recorded 26 PentAGI records on 25 hosts, mostly on ports 80, 8443, 443 and 3001, on CTG Server Limited, Hetzner and DigitalOcean networks.
- Neighbours. Monograph 19 found Cobalt Strike and PentAGI on the same host. Monograph 17 found a single PentAGI listener on port 8443 in one cloud range, next to DNSLog, CyberStrikeAI, ARTEX and Sliver listeners.
Most PentAGI installs are probably testers and hobbyists. Treat a PentAGI console as a hunting lead and look for a second sign before acting.
How to detect PentAGI
- Flow creation calls. PentAGI starts work through GraphQL requests to its API. Requests that create flows or spawn subagents from hosts outside your test team are worth review.
- Container parents that launch scanners. Alert when a container runtime process such as
containerd-shim-runc-v2ordockerdstarts nuclei, sqlmap or naabu with batch flags on a host with no testing role. - Certificates, carefully.
CN=pentagiorCN=localhoston ports 8443, 8666 or 8444 is a weak lead on its own. Pair it with an API match before acting. - Files on disk. PentAGI configs mention the GraphQL endpoint, pgvector, the Docker socket and the vxcontrol Kali image together.
Monograph 11 publishes a Suricata signature for PentAGI flow creation, a certificate hunting signature, a YARA rule for PentAGI configuration files and a Sigma rule for container-spawned scanners. All are in the Defense Kit.
How to govern PentAGI and defend against AI-driven testing
- Keep installs private. Leave the console bound to localhost or put it behind a VPN. Change the default admin password on day one.
- Scope every run. Require written scope and log every flow, as you would for a human tester. See the AI Agent Governance Policy Template.
- Protect model keys. A PentAGI server holds provider API keys. Store them in a secrets manager and rotate them if the console was ever exposed.
- Harden what agents will hit. Agents look for the same weak points as people. Patch internet-facing apps, and disable file upload paths that allow server-side script execution.
PentAGI compared with other AI offense frameworks
| Framework | How it is built (Monograph 11) | Common ports | Endpoints in Monograph 11 |
|---|---|---|---|
| PentAGI | Go controller, GraphQL API, Docker-in-Docker Kali workers, pgvector memory | 8443, 8666, 443 | 340 (22 confirmed) |
| HexStrike AI | Python server exposing nmap, nuclei, sqlmap and ffuf as MCP tools | 5000, 80, 8080, 8000, 8888 | 118 (0 confirmed) |
| CyberStrikeAI | Go and Python platform with role-based agents around 100+ tools | 5001 to 5003, 8080, 8082, 443, 8899 | 1,720 (1,290 confirmed) |
| ARTEX | Python multi-worker swarm with 1 to 16 subagents and a task queue | 8787, 443, 80, 7878 | 2,121 (495 confirmed) |
Frequently asked questions about PentAGI
What is PentAGI?
PentAGI is an open-source, self-hosted multi-agent system for autonomous penetration testing. AI agents plan a test, run security tools in Docker containers and write a report from a plain-language goal.
Is PentAGI malware?
No. PentAGI is an MIT-licensed security tool built for authorized testing. It is dual-use, because the same automation can be pointed at systems without permission.
Who made PentAGI?
PentAGI is published by the vxcontrol organization on GitHub, which also lists pentagi.com as the project website.
Is PentAGI legal to use?
Yes, on systems you own or are authorized to test. Scanning or exploiting other systems without written permission is illegal in most countries.
Which tools and AI models does PentAGI use?
The README lists more than 20 built-in tools, including nmap, Metasploit and sqlmap. It supports OpenAI, Anthropic, Gemini, AWS Bedrock, Mistral, other OpenAI-compatible endpoints and local models through Ollama or vLLM.
What are the PentAGI system requirements?
Docker and Docker Compose or Podman, at least 2 vCPU, 4 GB of RAM and 20 GB of disk, plus internet access. The web console runs on https://localhost:8443 by default.
How do I detect PentAGI on my network?
Watch for GraphQL flow creation calls, container runtimes that start scanners such as nuclei or sqlmap, and PentAGI configuration files. Odin Research publishes Suricata, YARA and Sigma rules for these signs.
PentAGI vs HexStrike AI: what is the difference?
PentAGI runs its own agent team and tool containers behind a web console. HexStrike AI is a Python server that exposes tools such as nmap and sqlmap as MCP tools, so an outside AI client drives them.
Sources
- vxcontrol, PentAGI repository, GitHub. github.com/vxcontrol/pentagi (checked 21 November 2026)
- vxcontrol, PentAGI README. raw.githubusercontent.com/vxcontrol/pentagi/master/README.md (checked 21 November 2026)
- MITRE ATT&CK, Container Administration Command (T1609). attack.mitre.org/techniques/T1609/ (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 HexStrike AI?
HexStrikeAI is a public project that exposes many security tools to AI agents through the Model Context Protocol (MCP).
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.