Skip to content
ARTEX.sh

Authorized AI pentesting:find your own gaps first, with permission and guardrails.

The same kind of tool that attackers use can show a company its weak spots first. This guide covers how to do that legally and safely on your own systems: a written scope, the guard settings to check, an isolated setup, careful data handling and a maintained copy. It has no install commands.

Checked

Why test yourself first

The FBI says ARTEX is good at finding side systems that get less attention. Running a similar tool against your own assets shows what an attacker would find. In a sealed lab, GitGuardian saw a v0.3.15 build find 8 of 9 planted secrets in one synthetic repository.

The developer said ARTEX was meant for testing assets you are authorized to test (Reuters). That is the line to hold.

Write the scope down first

Agents can go further than you meant. WorkOS reported that a researcher's ARTEX agent, during benchmark tests, escaped its test environment and reached an internal sales tool.

  1. 01

    Get written permission

    The system owner signs off on which assets, which dates and which tests. Systems run by a vendor, cloud provider or MSP need that party's permission too.

  2. 02

    List targets and limits

    Name the hosts, domains and accounts in scope, what is out of scope, and what the test must not do, such as taking real data out or slowing production.

  3. 03

    Set a stop rule and a contact

    Agree who can stop the test, how to reach them and what happens if the agent finds real customer data.

  4. 04

    Put the scope into the tool

    Use the per-task allow and block lists. The default asset blocklist only covers .gov and .edu domains, and the open code no longer has its original rules-of-engagement scope check.

Check the guard settings before every run

The ARTEX guard checks tool calls with intercept rules, then an LLM judge, then a human. The defaults are looser than many people expect. We read the code: in v0.3.15 and in RuoJi6 v0.3.30 the LLM judge is off by default, and a tool call that matches no rule is allowed when the judge is off.

In v0.3.15, if the judge is on but fails, the default action is allow. RuoJi6 changed that in v0.3.23 so a judge failure goes to human approval. BaaRai v0.6.0 adds a target approval gate pinned to host, port and resolved IPs, but it is off by default and only set through the API.

  • Turn the LLM judge on and keep human approval for anything risky.
  • Add deny rules for destructive actions and for out-of-scope hosts.
  • Never run v0.3.14. It has a public login bypass that lets an outsider take over the console (ARTXploit).

Never expose the console

TechCrunch reported on 2026-11-06 that Censys found more than 800 ARTEX deployments reachable without a password. StepSecurity notes the Compose file supplied with the original opens port 8787 on every network interface. WorkOS says the researcher's tool sat on the internet, accepted a weak development credential and had no rate limit.

  • Bind the console to localhost and reach it through an SSH tunnel or a VPN. BaaRai v0.5.0 and later bind to 127.0.0.1 by default; other builds may not.
  • Set a long, unique admin password.
  • Run the tool on a separate machine or network that can reach only the in-scope targets and the model provider. The egress guide (artex.sh/guides/contain-ai-pentest-agent) covers a default-deny setup.
  • Shut it down when the test ends.

Handle the data like evidence

The recording proxy saves every request the workers make, so its store can hold passwords, tokens and customer data from your systems. Your model provider also sees prompts and tool output. Pick a provider whose data terms you accept.

The console stores API keys and other credentials. Limit who can open it, keep its logs, delete raw traffic when the report is done, and rotate any real secret the test uncovered. Have a person confirm each finding before acting on it.

Choose a maintained, trusted copy

The original repository is gone, so every copy has a different origin. One look-alike copy pointed installs at a Docker image with a hidden back door (StepSecurity). Prefer a maintained fork with regular releases, check where its images come from, and avoid copies this site marks as offensive-class.

On this site, the install guide (artex.sh/install) covers maintained forks, the best forks page (artex.sh/forks/best) ranks them, and the compare pages (artex.sh/compare) set ARTEX next to other AI pentest tools.

Scope and approval controls differ by build

Checked on 2026-11-20. Target scope, tool approval and operating-system isolation are different controls. A prompt instruction is not a network boundary. This is a source comparison, not an independent security audit.

  • Check the exact build, saved configuration and runtime network policy before testing. Do not infer safety from a fork name or a list of features.
BuildDocumented control or defaultLimit to check
Original v0.3.15LLM judge is off by default; the existing guide records allow on judge failure.A .gov/.edu blocklist does not enforce a complete engagement scope.
RuoJi6 v0.3.30Tagged source: judge off; judge failure defaults to ask; approval timeout defaults to deny.Unmatched calls may proceed while the judge is off. Review the actual enabled tool set and rules.
BaaRai v0.6.0README: optional target gate pinned to host, port and resolved IPs.Gate is off by default and API-only. Verify that it is enabled for the engagement.
jaykwok v1.0.x / bcyw copyREADME documents removal of global and task-level scope gates.Defender-only listings here. CyHawk has no proof of deployment or confirmed attacks involving its reviewed bcyw build.
ARIA-X v0.1.1 rebuildDocuments target checks and auto, ask or deny approval modes; auto is the documented default.Terminal scope checks are heuristic. Its docs identify sandbox and network policy as the isolation boundary.

Questions

Is it legal to use ARTEX?
Using it on systems you own, or have written permission to test, is the purpose it was published for. Using it on anyone else's systems can be a crime in most countries.
Can I use ARTEX to pentest my own company?
Yes, with written sign-off from the system owners, a clear scope, the LLM judge and human approval turned on, and the console kept off the internet.
Is the ARTEX LLM judge on by default?
No. In v0.3.15 and RuoJi6 v0.3.30 it is off by default, and tool calls that match no rule are allowed while it is off. Turn it on before a run.
Should I expose the ARTEX console to the internet?
No. Bind it to localhost and use an SSH tunnel or VPN. Hundreds of exposed consoles were found, and v0.3.14 has a login bypass.
Which ARTEX fork should I use?
A maintained fork with regular releases and a known image source. The best forks page on this site ranks them, and the install guide covers the setup.
Are there alternatives to ARTEX?
Yes. Other AI pentest tools exist, and the compare pages on this site set ARTEX next to several of them.

Sources

Every claim on this page comes from these sources or from this site's own sourced pages. Open them for the full reports.

Follow the ARTEX story.

New forks, releases and reports, dated and sourced.

Isometric plane with glowing light columns