Yaamlabs

Threat intel

FBI FLASH: ARTEX AI is running whole intrusions

The FBI warns ARTEX, an AI attack framework, is automating intrusions in under three hours. What the advisory says and what to lock down now.

The Federal Bureau of Investigation (FBI) has issued a FLASH alert, reported as FLASH-20261113-001 and published around 13 November 2026, warning that criminals are using ARTEX, an artificial-intelligence (AI) penetration-testing framework, to automate intrusions from the first scan through to stealing data. FLASH is the format the FBI uses to push technical detail and indicators of compromise (IOCs) to industry. The advisory says the tool chains a large language model (LLM) over normal attack tooling so that one operator can run the whole sequence, and that some intrusions moved from initial targeting to privileged access in under three hours.

This is not a research demo. The FBI links ARTEX to real corporate intrusions, including an October 2026 supply-chain breach of ConnectWise Cloud Backup, formerly SkyKick Cloud Backup. The advisory and the research behind it describe more than 1,600 publicly exposed ARTEX consoles sitting on the internet in early November, with the real number likely higher because many run behind firewalls. The original ARTEX code was pulled from its public repository in October, but forks, archived copies, localised builds and a paid "ARTEX PRO" edition keep it in circulation, so a takedown alone does not close this off.

How it works

ARTEX is an agent framework, not a model. The advisory describes a program, written in Go, that does not ship its own intelligence and instead plugs into an external LLM such as Anthropic Claude, OpenAI GPT, DeepSeek, GLM, Grok or Kimi. The operator gives it a target. The agent then runs a loop: it reconnoitres, picks a weakness, writes and fires the exploit, escalates privileges and pulls out data, feeding each result back to the model and letting the model's answer choose the next move. The human effort is close to setup and supervision, which is why a single actor can keep several intrusions running at once.

An exposed ARTEX console is dangerous for two reasons. On your own network it is shadow tooling: an unauthenticated web interface that anyone who finds it can drive against you, or that leaks what an operator is already doing. On the attacker side, the exposed consoles are what let researchers count and study the campaigns in the first place.

The advisory is specific about where the agent is pointed. It calls out cloud administration surfaces, naming the Kudu/SCM endpoint on Microsoft Azure App Service. Kudu, reached through the Source Control Management (SCM) "advanced tools" site, is the deployment engine behind an App Service app, and it carries file-system and command access, so reaching it is close to reaching the application host. It describes abuse of OAuth (open authorization) application consent grants in multi-tenant applications: an attacker who compromises an application identity can reuse the permissions already granted to it to authenticate into downstream customer tenants, which is the mechanism behind the ConnectWise case. From there the agent pivots into backup and cloud-management planes. The "under three hours" figure matters because it compresses the window a defender has to notice and respond to something that, done by hand, used to take days.

What attackers are doing, per the advisory

The FBI says multiple actors have weaponised ARTEX since at least August 2026, against financial, government, technology, defence and healthcare targets among others. In the attacker-controlled instances that analysts examined, Chinese-developed models predominated, with the DeepSeek Flash family the most common; one write-up puts this at ten analysed instances, and the advisory is careful not to turn that into a general claim about any model's safety. The framework is described as most effective against the systems defenders tend to overlook: partner portals, internal employee applications, legacy services and third-party integrations.

On scale, the advisory points to more than 1,600 exposed consoles in early November. The internet-scanning firm Censys, whose work is cited in this reporting, published its own count of 1,465 exposed hosts running an ARTEX console as of 4 November, and noted that roughly 650 of those sat in two identical hosting blocks, so the exact total depends on how you count and when. Either way it is in the thousands and rising.

The ConnectWise Cloud Backup breach is the clearest example the advisory gives. Attackers compromised the application identities used by the backup service and used their standing permissions to reach customer tenants. Huntress reported at least 162 compromised tenants, and the attackers attempted to authenticate into more than 2,250 downstream Microsoft 365 tenants; Microsoft attributed the activity to the China-aligned actor it tracks as Storm-0940. The ARTEX repository was removed on 8 October after the abuse reports, which is why the forks and the paid edition matter: the code outlived its takedown.

What to do

Work through these in order. The first two close the surfaces the advisory names; the rest shorten the time you have to react.

  1. Hunt for and shut down any exposed ARTEX or similar AI-agent consoles on your own networks. Treat an unknown agent console the way you would an open admin panel: find it with your attack-surface scans, confirm who runs it, and take it off the internet.
  2. Lock down Azure App Service Kudu/SCM. In each app's configuration, set the SCM advanced-tools site's access restrictions to your admin IP ranges only, and turn off SCM and File Transfer Protocol (FTP) basic-authentication publishing credentials so the deployment endpoint cannot be reached with a static password.
  3. Review and reduce consent grants on multi-tenant applications, remove permissions no app still needs, and require administrator consent for new ones so a single compromised app identity cannot reach tenant after tenant.
  4. Restrict cloud-management tools and backup systems to trusted networks, and move administrators onto just-in-time (JIT) access, where a privileged role is granted for a short, logged window rather than held standing.
  5. Watch for the pattern, not just the IOCs. Alert on fast reconnaissance-to-privilege sequences, unusual bursts of automated portal lookups, unexpected OAuth consent grants, new Microsoft Entra ID credentials added to applications, and large outbound data transfers. The advisory notes its listed IP addresses may already have changed, so behaviour outlasts indicators here.
  6. Check your own suppliers. Ask your backup and remote monitoring and management (RMM) vendors whether they have addressed the application-identity exposure behind the ConnectWise case, and review your authentication logs for their application identities.

The wider lesson

What ARTEX changes is cost. The advisory describes work that skilled operators could always do, now done faster and with far less human time, which means more targets reached in parallel and a much shorter gap between the first probe and real access. The defences that hold up are the ones that do not depend on catching a slow attacker: least privilege on cloud administration surfaces, so a foothold does not become a tenant-wide compromise, and detection tuned to behaviour and speed rather than to yesterday's indicators.

Finding the forgotten portals and admin endpoints that this framework is built to hit is the first job, and it is where our attack surface management work starts; wiring up the fast, behaviour-based alerting the advisory calls for is what our security operations team does. Open the chat and Yaali, our AI agent, will pass your question to the engineer who would do the work.


Sources: FBI IC3 advisory (PDF), BleepingComputer, Neowin, GBHackers, eSecurity Planet, Censys exposure research.

Back to the blog, or read this post on the full site.