This is some text inside of a div block.
< back to blog

ARTEX, and the offensive tool powered by LLMjacking five months earlier

Crystal Morin
ARTEX, and the offensive tool powered by LLMjacking five months earlier
Published by:
Crystal Morin
Sr. Cybersecurity Strategist
Subscribe to Threat Research
ARTEX, and the offensive tool powered by LLMjacking five months earlier
Published:
November 20, 2026
Table of contents
TRT in your inbox

Get security briefings from the Threat Research Team

New campaigns, attacker techniques, and detection guidance — straight to your inbox.

Subscribe
falco feeds by sysdig

Falco Feeds extends the power of Falco by giving open source-focused companies access to expert-written rules that are continuously updated as new threats are discovered.

learn more
Green background with a circular icon on the left and three bullet points listing: Automatically detect threats, Eliminate rule maintenance, Stay compliant, with three black and white cursor arrows pointing at the text.

On June 12, 2026, the Sysdig Threat Research Team (TRT) watched someone build an autonomous hacking tool using someone else’s AI server. Then, in July, a similar tool called ARTEX appeared on GitHub. On November 13, the FBI released a FLASH saying a variety of threat actors have been running ARTEX against several business sectors. 

To be clear, there is no proof to date linking these two offensive tools. However, they are similar in nature, and defenders can learn from both about why they shouldn’t underestimate their speed, cost, or ability. 

The Sysdig TRT has a name for the kind of operator these tools enable: An agentic threat actor (ATA) is an operator whose attack capability is delivered via an AI agent rather than a human at the keyboard. ARTEX is the newest entry.

This blog covers what each tool did and what defenders can learn from putting them side by side. 

What the Sysdig Threat Research Team documented in June

Over the summer, the Sysdig TRT saw someone using a misconfigured Ollama server as the reasoning engine to build an automated, multi-stage offensive tool. The actor wasn't chatting with the model or reselling access as seen in prior LLMjacking operations. Instead, they wired it into a pipeline that scanned a target, matched it to known vulnerabilities, wrote proofs-of-concept (PoCs), and tried to break in, with the model making the decisions at every step. The tool self-identified throughout the code as VAPT, shorthand for vulnerability assessment and penetration testing, so we dubbed the tool VAPT.

Because VAPT sent its full instructions to the model on every request, we captured the whole architecture: fingerprinting, vulnerability matching, web recon, PoC synthesis, SQL injection, credential extraction, privilege escalation, and an orchestrator that ran until it got command execution. VAPT was being rewritten stage by stage while the Sysdig TRT watched, and every target was a private practice range. We were looking at a tool in development, before it was potentially turned on real victims.

Due to LLMjacking, the Ollama server was free, unauthenticated, and unmetered. For an actor whose tool makes hundreds of model calls per attack, our blog said, "the economics are clear." When the cost of an attack drops to the cost of inference, and inference is stolen, offensive operations become free.

What is ARTEX?

ARTEX is an open source, multi-agent penetration testing platform written in Go. It turns the operator’s goals written in natural language into machine commands run by specialized AI agents, with minimal human involvement necessary beyond setting the objective. 

ARTEX doesn’t have a model of its own. Microsoft describes its architecture as a planner agent to decide next steps and worker agents to run command tools, with all of the findings compiled in a shared graph of assets and vulnerabilities. 

Fortunately, the tactics, techniques, and procedures (TTPs) in the FBI FLASH read like an attack playbook every security team already knows: 

  • Sub-second permission checks and dry-runs
  • Exploitation of SAML, XXE, SSRF, injection flaws, and hardcoded credentials
  • Web shells and open source remote access trojans (RATs) for persistence
  • Kerberoasting and DCSync for escalation
  • Residential proxies, Tor, and tunneling tools used for command and control (C2)

The FBI says ARTEX traffic is noisy, which means defense evasion techniques aren’t an issue. It’s the speed that can pose a problem. Some intrusions moved from initial targeting to privileged access in under three hours, though according to Microsoft, most targeting attempts were unsuccessful. While other campaigns move faster with more success, three hours is nothing to brush off.

What VAPT has in common with ARTEX

The "authorized pentest" label is a master key. That framing recurred across VAPT's stages, and the operator likely used it as a wrapper to keep the model cooperative, even though the AI model was already abliterated, or had the guardrails removed. Whether VAPT was truly created for educational purposes or not, the compute engine used to build it was stolen. Therefore, the entire campaign was unauthorized. 

The FBI FLASH says attackers favor models with a "lack of observable guardrails when framed for penetration testing." Call the job a pentest, and many AI models stop asking questions. That's not one provider's bug, but a dual-use problem without a clean technical fix.

To that end, VAPT's first probes asked for an abliterated Llama-3.3-70B model. Its model list also included deepseek-r1:8b. The FBI says ARTEX operators gravitate to the DeepSeek Flash family because of its "balance of capabilities, relatively low cost, and lack of observable guardrails." Different models, but the selection criteria are the same: capable enough (they don’t need to be frontier models), cheap, and not likely to refuse a prompt.

VAPT's operator simply repointed a backend-agnostic tool at a free Ollama server in place of paid OpenAI, Anthropic, or Google APIs. ARTEX has no model of its own and plugs into at least six providers. In both cases, changing the engine is a configuration change. That's why "just block the bad model" isn't a strategy — it would require blocking all the models, and that just isn’t logical.

Another similarity is that neither VAPT nor ARTEX needs a novel exploit. There was no CVE in the VAPT operation, and the FLASH cites none for ARTEX either. The speed comes from chaining ordinary weaknesses without getting tired or pausing efforts. These AI-powered tools industrialize the boring stuff. And when pitted against human defenders, those who have to eat and sleep will always be at a disadvantage.

Lastly, both the tools and their fuel are sitting in the open. Researchers previously catalogued roughly 175,000 publicly exposed Ollama instances. And as of early November, researchers counted over 1,600 publicly exposed ARTEX consoles, according to the FBI. One is a free engine. The other is the thing the engine drives. The attackers have the same hygiene problem as the people they target.

Where VAPT and ARTEX differ

 

VAPT (June)

ARTEX (July onward)

Who

One actor on residential IP space in India

A variety of actors

What it is

One actor's tool, iterated within a session

An open source project with thousands of forks, translations, and a commercial edition

Targets

Private practice ranges

Real organizations across many sectors

Where the compute came from

Documented. An exposed and unmonitored server 

Not established

What defenders saw

The tool's full instructions, because the model server was one we could observe

The aftermath: bursts, dry-runs, new credentials, webhooks

The last row is particularly important. With VAPT, the Sysdig TRT saw the brain. With ARTEX, defenders mostly see what the brain did. Behavior-based detection exists for exactly this reason.

Cost is an important part of the story

Defenders can change many things, but the price of an attacker's tokens isn't one of them.

Sysdig TRT’s LLMjacking research has followed this cost curve since 2024. That May, we coined the term for attackers using stolen cloud credentials to reach a victim's paid AI models, leaving the victim with the bill, and modeled a worst case of up to $46,000 per day. By 2025, LLMjacking had matured into an industrialized black market. In June came the next step: exposed model capacity powering an offensive tool. The economics of compute theft can now fund the development of offensive capabilities.

One month later, ARTEX showed up, and the FBI's advisory reads like a footnote for LLMjacking:

  • Model choice followed price. DeepSeek Flash dominates attacker-controlled instances partly because of its "relatively low cost."
  • Resellers and gateways are in the picture. The FLASH's one domain indicator is an LLM API reseller used in the South Korean campaign, and it tells defenders to hunt for LLM gateway software (CLIProxyAPI, New API).
  • The FBI wants your AI bill. Among the evidence it asks victims for: token consumption patterns or API billing anomalies, unfamiliar LLM proxies or resellers in egress logs, and outbound connections to LLM APIs.

It’s important to note that it doesn’t say "stolen" credentials or access, because a reseller isn't proof of theft. However, LLMjacking is a natural fuel for tools like ARTEX, and the credential-theft half of one intrusion can supply the compute for the next. 

During a recent incident in October, the attacker used a harness similar to ARTEX and ran up a bill of $740,000 in unauthorized spend on stolen AI keys.

Separately, but further proving the point about what offensive security costs, Calif says it spent under $200 in tokens to find and build the first exploit for Java4Shell. 

The AI bill is telemetry that very few organizations pay attention to, or monitor in real time. An owner of an exposed model server sees elevated compute and an open port, not realizing it’s a multi-stage attack pipeline. A victim of stolen AI credentials sees an invoice. The FBI asking for token consumption in incident reports should be a pretty obvious hint about where defenders should be looking.

The challenge of “dual use” tools

Herein lies the struggle with many open source tools and attacks. ARTEX's developer said the tool was for authorized security testing within defined scopes. And the intent behind VAPT was unclear, because it could’ve been developed and used for legitimate activity, but stealing compute was illegal either way. 

The attack and the pentest are the same software, pointed at different targets. The "authorized" label works on the model, the repository gets pulled after abuse reports, and the forks keep going. The FBI says ARTEX still circulates through forks, archives, translations, and a commercial version. A GitHub query on November 18 returned around 1,100 repos, a few of which had thousands of forks. But blaming open source software isn’t the answer. And the transparency of open source software helps defenders to better understand the signals created by the tools themselves.

The defender’s playbook against autonomous hacking tools

Defenders worried about VAPT, ARTEX, and whatever nefarious autonomous offensive tool may come next can follow these recommendations. For ARTEX specifically, the FBI FLASH includes IOCs to hunt for as well. 

  • Treat AI infrastructure like a database. Don't leave model servers on public interfaces without authentication, and watch their usage. An exposed Ollama server is someone else's free engine.
  • Add the AI bill to your telemetry. Alert on token consumption spikes, unfamiliar API consumers, and billing anomalies, and know which credentials grant model access.
  • Treat AI-service credentials like cloud credentials. They're a high-priority target and a currency. Rotate and scope them, and keep them out of repos.
  • Run authorized agentic testing against yourself first, and treat any open-source pentest harness you run as hostile code: isolate it from production and enforce scope in configuration, not prompts.
  • Detect behavior, not models. You can't predict which model sits behind the next tool. What you can see is the pace: sub-second permission checks, bursts with small changes, sequential enumeration on partner portals, new identities probing other tenants, and servers reaching webhook.site or LLM APIs they never used before.
  • Vet indicators before you block. Cloud-hosted IPs can be reassigned, so make sure you verify the list and periodically review your block list. They don’t need to be permanently banned. 

Conclusion

In June, the Sysdig TRT watched someone build an offensive tool  on borrowed compute and aim it at practice targets. In November, the FBI described a polished, open source offensive tool in the field, run by a variety of actors against real organizations. There is no clear link between the two aside from the fact that neither required much money, skill, or novelty.

For two years, the question was whether AI agents could hack. That’s settled: they most definitely can. The question now is whether an organization will notice if they’re the one paying for it. 

About the author

featured resources

Test drive the right way to defend the cloud
with a security expert