With ARTEX, cybersecurity is starting to look like science fiction
AI can now break into systems autonomously.
What everyone needs to know, from executives to interns.
The cybersecurity world is currently buzzing about ARTEX. Much of the coverage is highly technical and, at times, alarmist. Yet this development affects everyone, not just cybersecurity experts. That is why we think it is important to explain, independently, what ARTEX is, what new era in cybersecurity it ushers in and what organizations can do to protect themselves.
By Patrick Coomans
On Friday, November 13, 2026, the FBI published a FLASH alert on ARTEX, a freely available software package that enables AI agents break into computer networks on their own. According to the FBI, in some cases attackers went from initial reconnaissance to obtaining administrator privileges in less than three hours.
In short. ARTEX is a freely available software package that lets AI agents break into corporate networks autonomously. According to the FBI, this has sometimes taken less than three hours and required little technical expertise from the attacker.
It is happening today, including Europe, and it concerns everyone in an organization. Attacks are getting cheaper and faster than most companies can keep up with. Companies that do not know what is exposed to the internet, or that cannot intervene quickly, face a greater risk of becoming the next cybersecurity headline, with all the direct and indirect consequences that entails.
ARTEX is both the trigger and the signal of a new trend that has now begun: variants are already circulating, and we can expect many more in the very near future.
Anyone who still thinks of AI only as a chatbot that sorts email or polishes a presentation is missing what is happening here. ARTEX is first and foremost a capability signal: evidence that an AI system can now turn an instruction given in plain language into a complete cyberattack, from reconnaissance to data theft, with barely any human involvement. That is why ARTEX belongs in the vocabulary of everyone who deals with cybersecurity, closely or from a distance, from the executive who approves the budget to the intern receiving a first account.
This is not a future scenario. The FBI warning is a few days old, and only a few weeks ago a major European e-commerce platform was attacked in this way for several days.
Whether your company already uses AI itself is not what determines this threat. In this case, it is the attackers who are using AI. Check Point found that most of the observed attacks were opportunistic and financially motivated: the attackers were looking for usable weak spots, without a single target chosen in advance. That meanssmaller or lesser-known organizations can also become targets.
A forgotten vendor access, for example, can be abused to get into a business application. Depending on the privileges attached to it, an attacker can steal customer data there, alter data or disrupt order processing. Organizations should therefore identify this week which critical business systems can be accessed externally, where the biggest risks and unknowns lie, and which measures need to be taken immediately.
Management should assign an owner, a budget and a deadline to that work.
1. What is ARTEX, and what can you do with it?
ARTEX is an open-source framework for penetration testing. A penetration test, or pentest, is an attack on request: a company has specialists try to break in to find the weak spots before a real attacker does. That work has traditionally been labor-intensive, expensive and dependent on scarce expertise.
ARTEX automates much of it. The user describes a goal in plain language, and the framework translates that objective into concrete commands that are executed by specialized AI agents.
An agent is AI software that goes far beyond giving answers the way a chatbot does: it creates a plan, launches hacking tools, interprets the results, adjusts its approach and tries again. Multiple agents can operate at the same time and share what they find with one another. According to CERT-EU, the cybersecurity service of the European institutions, a human operator can watch and steer through a dashboard.
The attacks described by Amazon, rely mainly on known techniques and security vulnerabilities. What AI changes is the speed and scale at which these techniques can be applied, across many targets simultaneously and without interruption.
As a result, a forgotten access or a missing security update can therefore be found and exploited sooner. Well-managed access, timely updates and active monitoring remain important and can limit or stop these attacks.
One important detail: ARTEX itself does not contain an AI model. The framework provides the orchestration and the hacking tools; the reasoning is done by a language model connected by the user. The FBI mentions both American models from Anthropic and OpenAI and Chinese ones such as DeepSeek, GLM and Kimi.
Just how little human work is still involved became clear when Check Point Research was able to look inside ten poorly secured ARTEX installations run by attackers. The attackers set a goal and after that mostly asked for a status update, literally: "What's the progress?" In one campaign, the agents exchanged 400,791 messages across 292 subgoals, without a single additional human instruction.
In the right hands, that is good news. An organization can now have its own systems tested continuously instead of only once a year, at a fraction of the cost. That was also the developer's intent:
authorized testing of predefined systems with the owner’s permission.
2. Can malicious actors use ARTEX too?
Yes, and they have been doing so since August 2026, barely a month after its release. Nothing in the software technically enforces t an agent to only test systems for which permission has been granted. Whether it stays within the specified targets depends on configurable rules and on the behavior of the connected language model; in CERT-EU's words, "there is no hard technical enforcement." Someone entering the address of another organization’s system can therefore use the same tool for an actual attack.
| Date | What happened? |
| July 2026 | Developer "Autumn-27" publishes ARTEX on GitHub, intended for authorized testing |
| August 2026 | First observed abuse by attackers |
| October 7, 2026 | CrowdStrike describes a campaign targeting South Korean financial institutions |
| October 8, 2026 | The developer takes the repository offline; copies keep circulating |
| October 21, 2026 | Microsoft reports a supply-chain attack at ConnectWise |
| November 5, 2026 | CERT-EU warns the European institutions |
| November 13, 2026 | The FBI publishes FLASH-20261113-001 |
Source: FBI FLASH and the reports cited in it.
Four months separate the release from the FBI alert. In that time, the FBI recorded attacks in financial services, government, technology, defense, semiconductors, energy, healthcare, manufacturing, telecom and the legal sector. At South Korean financial institutions, the personal data of about 67,000 people was stolen. Through a breach at ConnectWise Cloud Backup, attackers gained access to more than 2,250 customer environments. A single breach at a vendor thus opened the door to the systems of more than two thousand of its customers.
Security companies themselves are not spared either. Barracuda reported an ARTEX breach through an access key belonging to a former vendor whose credentials never had never been revoked, TechCrunch reports.
The Dutch firm Eye Security assisted a major European e-commerce platform that was under attack for several days. The attackers got in through a vulnerable upload function on the website and looked for things they could monetize, gift cards in particular. Three findings from the report show the pace. The attackers found an API key, a digital key that software uses to authenticate to a system, and were using it just minutes later.
When a limit on the number of requests slowed them down, the agents adapted their script in about three minutes to get around it. And less than an hour after the company had brought its systems back online, the attackers were back. CERT-EU considers it highly likely that ARTEX has also been used against European institutions.
3. Is ARTEX one of a kind?
No. Several comparable frameworks now exist, and ARTEX itself remains available even after the developer took it offline.
First, ARTEX itself . The developer took the code offline on October 8, but according to the FBI the framework remains in circulation through forks (copies that others continue to develop), translated variants, archived copies and even a commercial version, ARTEX PRO. In early November, more than 1,600 management consoles of ARTEX installations were publicly visible on the internet, and Check Point counted a total of 2,557 systems on which the framework was running or had run. CERT-EU warns that derived versions may have different capabilities than the original. In other words, the software can no longer simply be recalled.
Then there are the similar frameworks. On the same servers, Censys also found other AI attack frameworks, including CyberStrikeAI, PentAGI, HexStrike AI and RedAmon. The Google and Mandiant report cited by the FBI is titled "ARTEX and Friends," reflecting this wider ecosystem of related frameworks. New variants are to be expected because all the building blocks are freely available: software to orchestrate agents, existing hacking tools and a powerful language model of the user’s choice.
Safeguards introduced by AI model providers to prevent misuse offer only partial protection. Check Point observed attackers switch from an American model to Chinese models with fewer restrictions as soon as the American model began refusing their requests. The FBI notes that the DeepSeek Flash models appear most often, because of their capabilities, low cost, and apparent willingness to execute requests framed as penetration testing.
The ability to break in therefore sits in the language models themselves, and they get stronger every month. In July 2026, OpenAI disclosed that two of its models had escaped from a sandboxed environment during an internal test and had broken into the AI platform Hugging Face to get hold of the answers to a cybersecurity test, as Fortune reported. No ARTEX and no malicious human was involved in that incident. It demonstrated what language models may already be capable of on their own.
4. Why this keeps catching us off guard
Because even leading experts have for years systematically and badly underestimated how fast AI is improving. Wharton professor Ethan Mollick opens his new book Co-Existence with it: forecasters trying to estimate when AI will be able to solve top-level mathematics problems, and getting it badly wrong. In a conversation with Derek Thompson, he sums up the cause this way: "humans don't think in exponentials very well."
The figures come from the Forecasting Research Institute, which has been collecting forecasts about AI since 2022. It surveys a panel of 339 experts, including computer scientists, economists and people from the AI industry, alongside so-called superforecasters: generalists with a proven track record in forecasting. These are some results from the interim report of September 2026, as summarized by The Decoder:
| Question put to forecasters | What they expected | What actually happened |
| When will an AI system reach gold-medal level at the International Mathematical Olympiad, the annual competition for the world's best high school students? | The experts gave it an 8.6% chance of happening by 2025, the superforecasters 2.3%. The median expert did not expect it until around 2030. | An AI system reached that level in July 2025. |
| When will an AI system outperform a team of top virologists on a test about solving practical problems in a virology lab? | The experts expected this in 2030, the superforecasters in 2034. | That already happened in April 2025. |
| How much annualized revenue will the largest AI company reach by the end of 2026? | The experts estimated $20 billion, the economists $16 billion and the superforecasters $25 billion. | In September 2026, one company, Anthropic, was already at around $100 billion. |
Across all the questions about test results, the experts assigned on average a 24.6% probability to the outcome that later occurred. For the superforecasters, it was 9.7%.
Mollick draws an essential distinction between capability and adoption: what the models can do is evolving much faster than expected. What organizations do with them, the adoption, moves much more slowly.
For cybersecurity, that gap is the essence of the problem. An attacker can use a new AI tool right away: there is no procurement process, no change program, no employees to train and no legacy systems to take into account. A company that has to defend itself has all of those, and so adapts more slowly. Anyone who keeps adapting their security the traditional way and at today's mostly slow pace risks falling behind attackers who put new capabilities to work immediately.
That is bad news, and unfortunately there is no ready-made answer today.
Organizations have effectively had years in which they could continue digitalizing without always giving cybersecurity risks the attention they deserved. Getting hacked was seen as plain bad luck, an unlucky draw in the lottery.
Five lessons from strategic perspective
1.Attacks are happening faster and at a scale people cannot keep up with
When an attacker gets from initial reconnaissance to administrator privileges in three hours, a security team that only looks at alerts the next morning is too late. Arve Kjoelen, CISO of Barracuda, said after the breach at his own company that anyone who relies on detection and response alone "will often arrive after the attacker has finished."
Scale is equally important. An ARTEX operator does not have to divide attention between targets. Agents can simultaneously attack numerous systems and continue working around the clock.
At the European online retailer investigated by Eye Security, attackers targeted all available entry points at the same time, with peaks of more than 9,000 requests per minute sent from more than 17,000 different internet addresses to conceal their origin.
As a result, getting hacked is no longer an unlucky draw in the lottery. For an organization with a weak spot that is reachable from the internet, the time available to fix it can be very short.
2.An attack requires less expertise and less money. The FBI writes that ARTEX enables attackers with minimal technical knowledge to carry out complex attacks on multiple targets at once. The costs are low too. Check Point reconstructed a campaign of about two weeks against ten services that digitally sign iPhone apps outside the official App Store. The agents obtained more than 300 certificates, the digital credentials used to apply such a signature. The attacker’s estimated AI costs were just $700 to $1,500.
3.The techniques are known; they succeed where basic security is missing. CJ Moses, CISO of Amazon Integrated Security, writes that none of the observed techniques was new. The agents succeeded mainly where basic security was lacking: default passwords that had never been changed, passwords and keys sitting in source code, and management consoles reachable without a login. CERT-EU points out that such an agent also tests small configuration errors that a human tester usually leaves alone. Individually they are not very serious, but the agent combines them into one working attack. One vulnerability might reveal a username, another a password, and a third a system on which those credentials provide access.
4.Attackers get in through systems that receive little attention. According to the FBI, ARTEX is particularly effective against systems that are monitored less closely than the main business applications: portals for partners and vendors, servers behind mobile apps, applications for staff and outdated systems. Vendors are part of this as well. Anyone who gets into a less well-secured vendor can use that vendor's access to move on to the actual target. The unrevoked vendor key at Barracuda is one example.
5.The agents make mistakes and stand out. According to the FBI, ARTEX generates a lot of conspicuous network traffic, which makes an attack detectable.
The agents also get things wrong. The crypto campaign Check Point followed earned the attacker zero dollars, because the agents mistook a technical acknowledgment from the system for a successful payout. Amazon saw agents try to bypass an identity check with forged images (deepfakes); the check rejected every recorded attempt. Eye Security found that the attackers were clearly visible on the systems the company actively monitored. Those who watch their systems can therefore spot these attacks.
6. ARTEX plus the flood of vulnerabilities: the perfect storm?
Yes, the two reinforce each other. Far more vulnerabilities are being discovered, and attackers can exploit them faster and at greater scale.
First, the growing number of vulnerabilities. On November 13, the same day as the FBI alert, the number of registered software vulnerabilities (CVEs) for 2026 passed the 100,000 mark, more than double the total for all of 2025. To illustrate: Microsoft's monthly security update fixed 89 vulnerabilities in November 2024 and 63 in November 2025; in November 2026 the count was 1,094.
Josh Bressers of VulnCheck tells The Register that AI-driven bug hunting explains most of that increase. At the same time, the window for patching, that is, for installing security updates, is shrinking. The median time between the disclosure of a vulnerability and its first exploitation fell from 120 days in 2025 to 80 days in the first half of 2026. If public example code is circulating that shows how to exploit a vulnerability, then according to CrowdStrike 88% of the exploitation happens within 48 hours.
Now add an attacker that works day and night and tests many targets at once. So there are more vulnerabilities, less time to close them, and more attackers able to exploit them.
What changes fundamentally is the pace at which organizations have to patch. Anyone who does it once a month is too late when exploitation starts within 48 hours. Because no organization can patch everything at once, prioritization becomes important: address internet-facing systems and vulnerabilities that are already being actively exploited first, then work through the remainder, even if another vulnerability has a higher severity score on paper.
The U.S. government has required its own agencies to apply this type of prioritization since June.
7. How do we protect our companies and ourselves?
The answer is based mainly on familiar cybersecurity measures, but they need to be implemented more consistently and much faster than is often the case today.
The recommendations come down to five points. None of them requires new technology. What they do require is for management to assign clear ownership, budget, and deadlines
Know which systems are reachable from the internet.
Start today by creating an inventory of everything that can be accessed from the internet. Assign an owner and look beyond traditional IT.
Set a short deadline for identifying the most important access points, the biggest risks, and any remaining unknowns. If an inventory already exists, start by checking whether it is complete. If necessary, bring in external expertise.
Operational technology (OT) is just as much a part of it: machine controllers, sensors, remote maintenance access and connections with manufacturers and vendors, all the way to HVAC, cameras and physical access control. Do not forget the familiar blind spots either, such as vendor portals, 5G dongles on the production floor, old test environments and applications a business unit once set up on its own.
Remove systems from the internet that does not need to be there and put administration and maintenance access behind a shielded entry point with solid multifactor authentication. Also restrict where that access can be accessed from: allow connections only from the countries where your employees and vendors work or, more strictly, only from pre-approved internet addresses.
Manage passwords and digital keys very strictly. Systems and software authenticate to one another with digital keys, such as API keys and access tokens. Anyone who gets hold of such a key can pose as the legitimate user. Four measures limit that risk:
- Turn on multifactor authentication, where you provide a second confirmation in addition to a password. For example through an app on your phone. Start with administrator accounts and with every access from outside the corporate network.
- Give keys a limited lifetime and only the privileges needed for the task, so that a stolen key is usable only briefly and opens only a few systems.
- Revoke a vendor's access as soon as the relationship ends; at Barracuda, the attackers came in with a key from a vendor the company no longer worked with.
- Finally, check automatically that no passwords or keys have been left behind in source code and configuration files, because attackers use AI to search for exactly those.
Fix small security flaws too. In a security test or scan, every vulnerability found gets a severity score, and flaws with a low score often sit untouched for months. CERT-EU recommends fixing those as quickly as possible as well, because an AI agent combines several small flaws into one working attack. In addition, do not show detailed error messages to visitors who are not logged in. Such a message gives away, for example, which software and which version is running, and an agent uses that information to adjust its next attempt.
Make sure you can respond as fast as the attack unfolds. Decide in advance who is allowed, including at night and on weekends, to disconnect a system from the network and revoke access without further approval. Set thresholds and automatic alerts for applications that suddenly process far more requests than normal. Also watch for a sudden rise in the number of errors, such as failed logins or rejected requests.
An agent tries a lot and most attempts fail, which makes such a rise a very important indicator of an attack in progress. Test whether the logs from your systems still come in when you cut off the network during an incident, because without that data you cannot follow the attack. In its investigation, Eye Security itself used AI agents to analyze logs. Those agents were technically unable to change anything or read passwords, and humans made every decision.
Secure your own AI applications like any other system. Amazon saw attackers steal keys for paid AI services, leaving the victim with the bill for the AI usage of their next attacks. So apply the same access control and monitoring to AI applications as to other business systems, and have yourself alerted when usage rises unexpectedly.
And what can individuals do?
Three habits can make a significant difference.
- Turn on multifactor authentication or passkeys for your email, your bank and your social media.
- Use a password manager, so that every service has its own complex password.
- Let updates install automatically; never postpone them.
Beyond that, assume that your data will end up in a data breach sooner or later. That allows scammers to send messages containing accurate personal details. If you get an unexpected request to pay or log in, verify it through another channel, for example by calling a number you know yourself.
8. Looking ahead: things may get worse before they get better
It threatens to get much worse first, for three reasons. Attackers adopt new technology faster than defenders, while companies typically plan and budget their security by the year. The models are getting more powerful and cheaper, and not every provider builds in safeguards against abuse. And the number of new vulnerabilities keeps rising. Bressers expects that in a few years we will look back and ask, "remember when it was only 100K?"
In the longer term, however, the same technology could also improve cybersecurity.. ARTEX was built to test an organization’s own systems; they will soon be able to have that done daily instead of once a year. Software companies can use AI to track down flaws before they release their products. Security teams can have AI agents assess alerts in a few minutes. On top of that, European regulation such as NIS2 and the Cyber Resilience Act obliges organizations and software makers to improve their basic security.
All of that does require organizations to change how they work and to secure legacy systems, and that goes slowly. Several years can pass between the moment attackers master a technology and the moment most defenders have it in use. We do not know whether this period will last five years or fifteen, but that is the period we are in right now.
So do not wait for the next report or the next budget round. Put this on management's agenda this week and ask your IT lead or vendor for concrete answers: which risks have we identified, which ones are we limiting immediately and which decisions does that require from management? Also have them indicate where visibility into systems and digital keys is missing, and where rapid patching or intervention has not yet been arranged. Where necessary security or monitoring is missing, we have to assess which access we restrict or temporarily shut off, taking the consequences for business operations into account. Attach an owner and a date to the remaining measures.
If you still have questions about your company's cyber resilience, you can contact Kiwa via the form below.
Contact our expert
-
Patrick Coomans
- Global product owner cyber security
Contact us
Would you like to know more about our services in the field of Cyber Security? Please contact us using the form.