Managed ARTEX Pentest Service for Your Company
A managed ARTEX pentest is an ARTEX penetration test that we run for you. You give us the scope. We operate the agents, the recording proxy and the approval gates. You get findings that a human verified, with the HTTP request and response that prove each one.
ARTEX is a strong open-source engine. It has a planner, parallel workers, two linked graphs and a strict proof check. For the full technical description, read how ARTEX works. But a strong engine is not a pentest. Somebody must install it, configure it, watch it, approve its commands and check its results. This page explains what we do for you, and why it is easier and safer than a self-hosted run.
We are not affiliated with the ARTEX maintainers. ARTEX uses the AGPL-3.0 license.
Why ARTEX matters to your company now
ARTEX is no longer only a tool for authorized testers. Attackers use it against real companies. This changes the question. It is not “should we use an AI pentest tool?”. It is “what does this tool find when an attacker points it at us?”.
The public reports give this timeline:
| Date (2026) | Event | Source |
|---|---|---|
| 26 July | ARTEX appears on GitHub as Autumn-27/ARTEX | StepSecurity |
| August | First reports of malicious use | StepSecurity, ThreatDefence |
| 8 October | The author closes the source. The original repository goes offline. Forks stay public. | Cloudflare |
| 5 November | Docker removes a backdoored ARTEX image | StepSecurity |
| 12 November | ARTEX and its forks are over 60% of the AI pentest orchestrators that Cloudflare tracks | Cloudflare |
| 13 November | The FBI publishes a FLASH alert on ARTEX | ThreatDefence |
Cloudflare’s numbers show the scale. In three days, it saw more than 50 families of AI pentest frameworks scan its customers. Daily requests from these tools increased by 30 times after 1 October. ARTEX traffic increased by more than 10 times after the original repository went offline. More than 12,000 repositories contain code from the original project. These are Cloudflare’s own measurements. Nobody has verified them independently.
How the attacks start
The reports agree on one point: the attacks usually do not start with a new exploit. They start with a credential that somebody left open.
In one Cloudflare incident, attackers took a static CI/CD token that had no expiry date. They used it to move from a cloud provider into the Cloudflare account of the customer. There, ARTEX used an account-level API token with no IP restriction and no rate limit. It made many new Worker scripts with names such as artex-pgrecon-*. In the same report, Cloudflare saw 400 database dump requests in one hour against a customer.
StepSecurity reports a similar pattern: leaked secrets are the fastest way in for ARTEX operators. In another case, an agent logged in to an internal tool with a weak password that was for local development only.
An agent does not get tired. It tries each leaked key, each default password and each endpoint with no rate limit. A human attacker can miss one of these items. An agent probably does not.
What a managed ARTEX pentest shows you
A managed ARTEX pentest uses the same class of tool that attackers use, with the same planner and the same scanners. But we run it with written scope, human approval and human verification. Thus, you see the result of an ARTEX attack on your application before an attacker sees it.
We also look for the items that the reports name as the first step of real attacks:
- API keys and tokens in your JavaScript bundle, source maps and public files.
- Endpoints with no rate limit, for example login and password reset.
- Endpoints that accept requests without a session, or with the session of a different role.
- Development credentials and debug routes on production hosts.
What you must do to run ARTEX yourself
ARTEX is one Go binary, but a real test needs many more parts. Before the first useful result, your team must:
- Select a fork that you can trust, and review its code. The original repository is offline, and hundreds of forks have their own changes.
- Build the Docker image from that source. Do not pull an image from Docker Hub. Researchers found backdoored ARTEX images there, and StepSecurity reports approximately 9,000 pulls of one of them.
- Install and operate PostgreSQL.
- Get LLM keys and configure provider profiles, base URLs and failover.
- Trust the proxy certificate authority on each tool that sends traffic.
- Keep the console off the internet. Put it behind a VPN or an SSH tunnel, or behind a reverse proxy with HTTPS. Turn off SSE buffering. Internet scans found more than 1,000 exposed ARTEX consoles in November 2026.
- Set
ARTEX_TRUSTED_PROXIES, or one attacker can lock all users out of the login. - Back up the database and the
keys/folder separately. Without the key, the backup cannot restore the secrets. - Write the scope rows, the intercept rules, the noise profile and the budgets.
- Stay at the console during the test to approve tool calls.
- Tell your SOC and your WAF provider about the test. Security vendors now publish detections for ARTEX traffic, so your own controls can block the test or start an incident.
- Read each finding and check that the proof is real.
Each step can go wrong. A backdoored image gives an attacker your LLM keys, your target list and your findings. A wrong scope row sends the agents to a system that you do not own. A missing rate limit can overload your production API. A wrong budget can stop the test before the first worker starts, or let it spend a large amount on tokens.
What we do for you
With the managed service, you do none of the steps above. Our engineers do them. The table shows the division of work.
| Task | Self-hosted ARTEX | Managed ARTEX pentest |
|---|---|---|
| Infrastructure, database, backups | Your team | We do it |
| LLM keys, providers, cost | Your team | We do it |
| Proxy CA and traffic capture | Your team | We do it |
| Fork selection and code review | Your team | We build from reviewed source at a fixed commit |
| Console exposure | Your team | Our console is not on the public internet |
| Coordination with your SOC and WAF | Your team | We give you our source IPs and test window |
| Scope and rules of engagement | Your team | We write them with you, in writing |
| Approval of risky commands | Your team, at the console | Our engineer |
| Finding verification | Your team | Our engineer, by hand |
| Report for auditors and buyers | You write it | We write it and sign it |
| Retest after the fix | You run it again | Included for 30 days |
You give us three items:
- The scope: domains, applications and the systems that are out of scope.
- Test accounts: one account for each role that you want us to test.
- The rules of engagement: test windows, rate limits and an abort contact.
Then we run the test and send you the report.
How a managed engagement works
The procedure has five steps.
Step 1: Scoping call
We talk for 30 minutes. We learn what your application does, who needs the report and what your deadline is. After the call, you get a fixed price and a start date.
Step 2: Written rules of engagement
We write the scope and the rules of engagement. You approve them before we send one request. The document lists:
- The hosts and paths that we can test.
- The hosts and paths that we must not touch.
- The maximum request rate.
- The payload types that are not permitted, for example data deletion.
- The test windows and the blackout windows.
- The abort contact and the abort procedure.
We put these rules into the ARTEX scope rows, the target gate and the intercept rules. The agents cannot explore assets outside the scope.
Step 3: The orchestrated run
We run the ARTEX planner and workers against your scope. The planner divides the goal into intents. The workers run real tools, for example nuclei, ffuf, sqlmap and curl. All HTTP traffic goes through the recording proxy.
During the run, our engineer:
- Approves or rejects each risky tool call. ARTEX denies destructive calls automatically, and a human approves the calls that the rules do not cover.
- Watches the request rate and the noise counters for each host.
- Stops a worker that repeats work without a result.
- Steers the planner to the business logic that agents usually miss, for example multi-role access paths.
For authenticated tests, we log in with your test accounts through the capture extension. This includes CAPTCHA, email OTP and 2FA.
Step 4: Human verification
ARTEX marks a finding as verified only when exact quotes from a recorded response support it. This removes many false findings. But we do one more check. Our engineer reproduces each finding by hand against your application. The engineer also checks that the severity is correct. If a finding does not reproduce, we remove it.
Step 5: Report and retest
You get a report with:
- The scope and the methodology.
- The findings, sorted by severity.
- For each finding: the request, the response and a cURL command that you can run.
- Remediation guidance for your stack, plus a prompt that you can paste into Claude Code or Cursor.
- The name of the engineer who ran the test.
When you fix the findings, tell us. We run the test again on these findings in the 30-day retest window. Then you get a letter that you can send to your customer or auditor.
Advantages of a managed ARTEX pentest
No setup and no operation
You do not install a database, a proxy or a binary. You do not manage LLM keys or token budgets. You do not keep a server, backups or encryption keys. Your engineers continue to build your product.
Less operational risk
An autonomous agent with sqlmap is a risk to a live system. Our engineers know which ARTEX controls are hard limits and which controls are only advisory. For example, the ARTEX noise governor does not slow down raw scanners such as nmap. The prompt-injection markers decrease the risk, but they do not stop a model from following injected text. We add human approval and written rules on top of these controls.
Findings that a human confirmed
Agents are fast, but sometimes they are wrong with high confidence. ARTEX has good proof checks: proof-of-exploitation levels, quote checks and a retester. We add a human who reproduces each finding. You get no “possible SQL injection” items in your report.
A report that your buyer accepts
A self-hosted ARTEX run gives you findings and an export. It does not give you a signed document with a methodology section. Enterprise buyers, SOC 2 auditors, ISO 27001 assessors and cyber insurers ask for that document. We write it and sign it. For the details of each framework, read compliance penetration testing.
No supply-chain risk from forks
ARTEX now exists as hundreds of forks, and some public images contain a backdoor. A self-hosted run makes you responsible for the code that you execute. We build from source that we reviewed, at a fixed commit. We do not use images from public registries.
Your SOC sees a test, not an incident
Security vendors now publish detection rules for ARTEX traffic, for example its user agents and request patterns. If your SOC sees ARTEX traffic with no warning, it starts an incident. Before we start, we give you our source IP addresses, the test window and an abort contact. You can select one of two options:
- Allowlist the test. Your WAF lets our traffic through, and we test the application itself.
- Keep your controls on. We test the application and your detection together. The report then also shows which ARTEX actions your WAF, rate limits and bot controls stopped, and which actions they did not stop.
The second option answers a question that a normal pentest does not: does your defense detect an AI pentest orchestrator?
Fixed price, no surprise token bill
A self-hosted run uses your LLM account. A long test with many workers can use a large number of tokens. With the managed service, you pay one fixed price per engagement. The model cost is our cost, not yours.
Expert steering
The ARTEX planner decides what to test next. A weak plan gives weak results. Our engineers steer the planner to the high-value paths: access control between roles, tenant isolation, payment flows and business logic chains. These are the paths that an automatic planner often skips.
Your data stays controlled
We record the traffic for the evidence only. We keep it encrypted at rest. We delete it after the retest window, or at the date that you select. Tell us your data requirements on the scoping call.
Self-hosted or managed: how to decide
Run ARTEX yourself if:
- You have a red team that wants full control of the prompts, skills and models.
- Your rules do not permit test traffic from a third party.
- You want to use a local model through Ollama.
- You have time to operate the infrastructure.
Use the managed service if:
- A customer, auditor or insurer asked you for a pentest report.
- You do not have a security engineer to watch the agents.
- You want a fixed price and a fixed date.
- You want a human to confirm each finding before you see it.
Many teams use both: a managed engagement for the report, and continuous vibe-coding security scans on each deploy.
How to defend against ARTEX between tests
A pentest is a check at one point in time. Do these steps also. They come from the Cloudflare and StepSecurity reports:
- Find secrets in your repositories and in the git history. Rotate each secret that you find.
- Replace static CI/CD tokens with short-lived OIDC credentials. Set an expiry date on each personal access token.
- Limit each API token to the IP addresses and the permissions that it needs.
- Set rate limits on login, password reset and other sensitive endpoints. Set them per account and per IP address.
- Validate each API request against an OpenAPI schema. Add a default-deny rule at the end of each rule set.
- Put internal tools behind identity-based access, not only behind a password.
- Use phishing-resistant MFA, for example FIDO2 keys, for admin accounts.
Between pentests, run continuous security scans on each deploy, so that a new leaked key or open endpoint does not stay open for months.
Get started
Book a 30-minute scoping call. We discuss your application, the people who need the report and your deadline. You get a fixed quote and a start date. See the pentest packages and prices.
Common questions
What is a managed ARTEX pentest?
Do I need to install ARTEX?
Are you the ARTEX project?
Will the agents damage my production system?
Is the report accepted by auditors?
Why test with ARTEX if attackers use it?
Will my WAF or SOC block the test?
Is it safe to download an ARTEX Docker image?
What does a managed ARTEX pentest cost?
Let us run the pentest
Tell us the scope. We operate the agents, approve each risky command and verify each finding by hand. You get a signed report and a 30-day retest.
14-day free trial · Cancel anytime