ARTEX vs PentAGI:two self-hosted Go systems with different guards.
PentAGI is an open-source multi-agent pentest system from VXControl, first published in January 2025. ARTEX is a 2026 platform that now lives on in forks. Both are written in Go and run on your own servers. This page compares design, license, safety controls, status, published numbers and attack reporting, as of 2026-11-17.
Data as of 2026-11-17
ARTEX and PentAGI side by side
| Aspect | ARTEX | PentAGI |
|---|---|---|
| Maintainer | Autumn-27 wrote it. The original repo was deleted on 2026-10-08. RuoJi6/ARTEX is the main maintained fork | VXControl, a security engineering company in Dubai |
| Language and stack | Go 1.26, Next.js 16 console, PostgreSQL 16 | Go backend with REST and GraphQL APIs, React console, PostgreSQL with pgvector, optional Neo4j graph |
| License | AGPL-3.0 | MIT |
| Architecture | A planner and 3 workers by default over an asset graph and an exploration graph | A team of specialist agents (research, coding, infrastructure, reporting) with long-term memory, optional planning and execution monitoring |
| Guard and safety controls | Every tool call passes intercept rules, an LLM judge and human approval. A recording proxy keeps all traffic. Default block on .gov and .edu | Tools run in an isolated Docker sandbox. Hard tool call limits per agent (100 or 20). The README describes no approval step per tool call |
| Deployment | One Go binary with a web console and PostgreSQL, from release zips or Docker | Docker Compose, set up by an installer for Linux, macOS and Windows. Optional Langfuse, Grafana and Neo4j stacks |
| Maintenance | RuoJi6/ARTEX: v0.3.30 on 2026-11-14, last push 2026-11-14, 2,194 stars (2026-11-17) | v2.2.0 on 2026-10-05, last push 2026-11-13, 34,440 stars and 5,454 forks (2026-11-17) |
| Offsec Harness Index | Entry 01, tier Confirmed, breach stated. 14 primary reports. Key figure: 200+ organizations targeted (CrowdStrike) | Tier Confirmed, breach stated. 4 primary reports, from Anthropic, Google (two) and CERT-In. Key figure: 12 to 26 GB of databases taken |
| Public benchmark | Baidu Agent+ final: 1st overall (2026-09-03). Live range 67.83/100, 7th. TSecBench 89.78/100, 6th | No benchmark score. The README claims 2x better results with small models when planning and monitoring are on, from its own tests |
Two Go systems, two shapes
PentAGI's README describes it as a self-hosted multi-agent system: you give a target and a goal in plain language, a team of specialist agents plans the work, runs more than 20 security tools inside an isolated Docker sandbox, keeps what it learned in PostgreSQL with pgvector and writes a report. An optional Graphiti and Neo4j graph tracks hosts, services and findings. An enterprise edition adds agent profiles, plugins and MCP servers.
ARTEX is also a Go service with a web console and PostgreSQL, but it is built around two graphs: a shared asset graph and one exploration graph per task. A single planner is the only agent that creates work; workers each claim one intent, run real tools, write facts back and stop. Both projects let you pick your own model. PentAGI lists more than 10 providers plus local models; ARTEX takes any endpoint in the Anthropic or OpenAI formats.
Safety controls
The two projects put their controls in different places. ARTEX checks each tool call before it runs: intercept rules first, then an LLM judge, then a person who can approve or reject it. All worker traffic goes through a recording proxy, so there is a record of what was sent. This site's data also notes that an earlier scope check was removed from the open code.
PentAGI isolates where tools run rather than gating each call. Its README describes a Docker sandbox for all tool runs, hard limits on tool calls per agent, a reflector that steers stuck agents toward done or ask steps, and optional monitoring and planning that are off by default. The README also warns operators that each flow sandbox publishes two TCP ports on every network interface by default, and tells them to restrict that range with a firewall. Defenders who find PentAGI on their own network should check that range.
In the Offsec Harness Index
The Offsec Harness Index lists ARTEX as entry 01 with the tier Confirmed, which means a vendor, response firm, CERT, agency or AI lab has shown it in a real operation. The index notes the original repo was closed in October 2026 and forks live on. It cites 14 primary reports for ARTEX; its key figure, 200+ organizations targeted, comes from CrowdStrike's report of 2026-10-29.
PentAGI is also Confirmed. Anthropic's report of 2026-09-10 named it as a framework used by two groups, one that stole production AI API keys and one that dumped political databases. Google's report of 2026-10-30 said most PentAGI activity it saw stayed at asset mapping or scanning, while some reached intrusion and persistence at several organizations. CERT-In's advisory of 2026-11-04 lists it in the CRAFTY FALCON campaign. These rows say what was reported. They are not a ranking, and they say nothing about the maintainers' intent.
Status, license and numbers
PentAGI is MIT licensed, which lets anyone reuse the code with few conditions. ARTEX is AGPL-3.0, so a modified version run as a network service must publish its source. PentAGI is the larger project by stars, 34,440 against 2,194 for RuoJi6/ARTEX on 2026-11-17, and it has a single maintainer with a steady release line.
ARTEX has public contest scores from the Baidu Agent+ final and TSecBench. PentAGI publishes no benchmark score. Its README says its planning and monitoring options gave 2x better results with a 27B open model in the project's own tests, with no method or data linked, so this page treats that as a claim.
Sources
- vxcontrol/pentagi on GitHub
- PentAGI releases
- Offsec Harness Index, PentAGI row
- Offsec Harness Index, ARTEX row
- Anthropic threat report, 2026-09-10
- Google Threat Intelligence Group, 2026-10-30
- CERT-In advisory, 2026-11-04
- CrowdStrike, 2026-10-29
- RuoJi6/ARTEX on GitHub
- Baidu Agent+ and TSecBench scores, cn-sec
Related on this site
More comparisons
CompareAntiFox vs ARTEX
AntiFox is a public fork that says it was built on ARTEX and adds report delivery. What changed, how active it is and its license.
ARTEX vs PentestGPT
A Go platform with a web console against a Python command-line agent from a USENIX Security 2024 paper.
ARTEX vs Strix
A self-hosted Go platform against an Apache-2.0 Python agent with a Docker sandbox, a cloud edition and CI support.
ARTEX vs CAI
An AGPL platform still maintained in forks against CAI from Alias Robotics, now archived with a commercial successor.
ARTEX vs HexStrike AI
A self-contained agent platform against an MCP server that lets your own AI client call more than 150 security tools.
ARTEX vs Cairn
Cairn is the one project the ARTEX README lists as its reference. Both grow a graph of facts and intents; one is a Go platform, the other a Python engine.
ARTEX vs CyberStrikeAI
Two Go platforms with a web console and human approval. CyberStrikeAI is Apache-2.0 and also ships WebShell management and a built-in C2.
ARTEX vs Shannon
Both AGPL-3.0. ARTEX runs open-ended assessments from a console; Shannon reads an app's source and only reports what it can prove.
ARTEX vs T3MP3ST
Both AGPL-3.0. ARTEX gates every tool call; T3MP3ST drives the coding agent you already use and has been quiet since 2026-09-08.
ARTEX vs AgenticPentest
ARTEX is AGPL-3.0 and gates every tool call; AgenticPentest is a two-week-old MIT workbench that fences work by scope and approves state-changing actions.
ARTEX vs ARIA-X
ARTEX and an independent Python rebuild: shared design ideas, different storage and control defaults, no performance ranking.
Follow the ARTEX story.
New forks, releases and reports, dated and sourced.
