SamoTech

TECHNOLOGY • APPS • IDEAS

SamoTech / ARTEX Explained: How an Open-Source AI Pentest Framework Became an Attack Tool and What Defenders Should Check

ARTEX Explained: How an Open-Source AI Pentest Framework Became an Attack Tool and What Defenders Should Check

Cybersecurity and AI news

A security testing tool built to help people find holes in their own systems is now, according to the FBI, helping criminals and state-backed groups break into other people's. The tool is called ARTEX. The more useful question is not "how scary is it?" but "what does it change, and what still stops it?"

What ARTEX is

Per BleepingComputer, ARTEX is an autonomous penetration testing framework written in Go. It does not ship its own AI model. It connects to outside large language models, and the FBI advisory, as reported by The Record and BleepingComputer, lists Claude, GPT, DeepSeek, GLM, Grok and Kimi among the models it can use. A user gives it a goal in plain language. The framework turns that goal into commands that AI agents carry out.

The reports differ on when it first appeared on GitHub (BleepingComputer says July 2026, The Record says August). Both say the developer removed the repository on October 8 after reports of abuse, and the FBI says copies live on through forks, archives, localized versions and a commercial edition called ARTEX PRO.

How an AI-driven intrusion is organized

Amazon's threat intelligence team had an unusual view. Its MadPot decoy sensors posed as popular LLM chat interfaces, and ARTEX agents sent requests to them as if they were real model backends. Because agent frameworks send their whole working history to the model, those requests carried each campaign's instructions, tool history and findings.

From that, Amazon describes a division of labor:

  • Planner, worker, verification and compression agents split the job.
  • A shared knowledge graph acts as the operation's memory. Every agent reads it and writes results back, so each one picks up where another stopped.
  • Standing instructions from a human set the mission, targeting rules and limits. In one campaign they told agents to work through countries in a fixed order and to treat credentials exposed for more than 30 days as low value.

Two campaigns Amazon described

Stealing AI access. Agents hunted for API keys to AI services. Amazon says stolen keys were fed back into the framework's own model gateway, so the stolen keys paid for the AI that searched for more keys. Targets were internet-exposed AI applications and gateways that store provider keys, because one weak admin page can expose many accounts.

Defeating identity checks. Agents tried to get past know-your-customer verification at two cryptocurrency exchanges using real-time deepfake video and altered ID photos. Amazon notes the vendor's server-side check rejected the deepfake in every recorded attempt, so this campaign is evidence of effort, not proof of success.

How this compares with older attacks

Amazon says none of the techniques were new. What changed is speed, persistence and the skill needed. The FBI says some intrusions went from first targeting to privileged access in under three hours, and that the framework lets people with little technical skill run complex, multi-target operations at the same time. In other words, the old playbook, run by software that does not get tired.

The FBI adds that ARTEX is good at finding places that get less attention: partner and supplier portals, mobile backends, employee apps, legacy services and third-party integrations. Reported cases include the ConnectWise Cloud Backup supply-chain attack, where, per BleepingComputer, attackers abused application identities to try to reach more than 2,250 downstream Microsoft 365 tenants, and attacks on South Korean banks.

What defenders can do

The advice in the sources is plain, which is the point. Amazon says the agents succeeded where basics were missing: default or hardcoded credentials, keys left in code and public files, unauthenticated interfaces and unpatched software. The practical checklist that follows from the reports:

  • Search your own code, container images and public repositories for exposed keys, and rotate any you find.
  • Treat AI apps, model gateways and inference endpoints like any production system: authentication on, defaults changed, patches applied.
  • Watch for unexpected AI spending or unusual usage, which can signal stolen keys.
  • Restrict access to cloud admin pages, review application permissions and authentication logs, and rate-limit sensitive APIs (the FBI's guidance as reported by BleepingComputer).
  • Expect noise: BleepingComputer notes these attacks generate many suspicious requests but move faster than manual response. Automated alerts and quick rotation matter more than reading every log line.

What is not known

Amazon assesses with moderate confidence that ARTEX was used and that the actors are separate Chinese-nexus groups, and it says it cannot verify the outcomes the agents reported. The true number of deployments is unknown: researchers counted over 1,600 exposed consoles, and the FBI says the real figure is likely higher. This article does not cover the technical indicators published by Amazon and the FBI. Defenders should read those directly, and treat IP-based indicators as possibly out of date, as BleepingComputer notes the FBI warns.

Sources: Amazon Threat Intelligence (CJ Moses), "Amazon Threat Intelligence observes threat actors running cybercrime campaigns with ARTEX AI agents" (AWS Security Blog, November 13, 2026, aws.amazon.com/blogs/security); The Record, "FBI flash alert says Artex tool automating cyberattacks at 'machine speed'" (November 13, 2026, therecord.media); BleepingComputer, "FBI warns ARTEX AI hacking tool is powering global cyberattacks" (November 15, 2026, bleepingcomputer.com). Read November 16, 2026. This is a summary and analysis of those reports. We did not review the FBI advisory itself.

Join the discussion on Blogger