Offensive AI agents such as ARTEX: Cybersecurity homework that now really has to be done, and done right

 · 15 min read
Contents6 sections
SeriesCybersecurity homework against AI attackersPart 1
  1. 1Offensive AI agents such as ARTEX

Since late August 2026, criminal hackers and suspected espionage groups have been breaking into companies and government agencies with the AI pentest framework ARTEX and its relatives; week after week, new cases come to light. The reports rarely describe sophisticated hacking: the AI agents dig up keys left lying around, find forgotten servers, and get in through unpatched software vulnerabilities, and they do it faster than humans. What helps is a list of IT security basics that have been known for decades and that (for now) need no AI.

From GitHub repo to FBI warning#

On July 26, 2026, a developer going by Autumn-27 posted the ARTEX framework on the code-hosting platform GitHub, open source and free of charge.Note 1 A planning agent breaks the human attacker’s goal down into subtasks; worker agents then launch established hacking tools such as the port scanner Nmap and evaluate what those tools return. ARTEX works with any large language model (LLM) that has an OpenAI-compatible API. Within a few weeks, the project collected thousands of GitHub stars. By late August, the first criminals were already using it for real break-ins. The ARTEX developer pulled the repository on October 8, one day after CrowdStrike’s report on break-ins at South Korean financial firms (see Figure 1). The forks live on, among them renamed copies and a commercial edition called “ARTEX PRO.” Internet scans turn up anywhere from hundreds to well over a thousand exposed web consoles.Note 2 On November 13, the FBI (PDF, opens in a new tab) issued its own alert about ARTEX.

Figure Less than four months passed between the release of the AI attack framework ARTEX and the FBI warning.
ARTEX timeline: 2026-07-26 ARTEX is released as open source on GitHub; late Aug 2026 Mandiant observes first attackers using ARTEX; 2026-08-31 Seoul suspect starts using ARTEX, police say; 2026-09-21 DIVD breach, an ARTEX agent takes over; 2026-10-07 CrowdStrike: breaches at South Korean financial firms; 2026-10-08 The author pulls the repository, forks remain; 2026-10-21 Microsoft: Storm-0940 via a ConnectWise app; 2026-10-29 CrowdStrike: VAPOR PANDA, more than 200 organizations; 2026-11-13 FBI warns in FLASH-20261113-001.2026-07-26ARTEX is released as open source on GitHublate Aug 2026Mandiant observes first attackers using ARTEX2026-08-31Seoul suspect starts using ARTEX, police say2026-09-21DIVD breach, an ARTEX agent takes over2026-10-07CrowdStrike: breaches at South Korean financial firms2026-10-08The author pulls the repository, forks remain2026-10-21Microsoft: Storm-0940 via a ConnectWise app2026-10-29CrowdStrike: VAPOR PANDA, more than 200 organizations2026-11-13FBI warns in FLASH-20261113-001

In the chronology of AI attacks that I maintain, ARTEX shows up in 40 entries as of November 15, 2026. Almost all of them are directed-autonomous, and the number of such cases has all but exploded since the start of the year (see Figure 2).

Figure In all eleven hacking cases that came to light in January 2026, AI played only a supporting role; in October, agents carried out steps themselves in 96 of 113 cases. Interactive in the chronology of AI attacks.
Published AI attacks per month in 2026; filled: cases in which agents carried out steps themselves: January 0 of 11; February 3 of 15; March 5 of 24; April 3 of 14; May 9 of 31; June 13 of 35; July 21 of 51; August 27 of 45; September 40 of 78; October 96 of 113; November 62 of 79 (November through 2026-11-15).Agents carry out steps themselvesAI only as an aid11Jan15Feb24Mar14Apr31May35Jun51Jul45Aug78Sep113Oct79NovNovember through 2026-11-15

In September 2026, I traced the attack by OpenAI agents on Hugging Face for the German IT magazine iX. My conclusion at the time, translated from German:

“Security leaders have to reckon with this state of the art: a human decides what gets hacked, and agents support them or take over individual steps or many, semi-autonomously or fully autonomously. AI augments every kind of attacker: Those who already hack well get faster and better; for those who can’t, the barrier to entry drops. All of them can do more damage, in breadth and in depth, than before. Defenders who don’t prepare for this will soon hit reality hard. In the medium term, any defender who does without AI, for whatever reason, widens the asymmetry even further and is unlikely to come through that unscathed.”

OpenAI-Agenten greifen Hugging Face an (in German) (PDF, opens in a new tab), iX 10/2026, p. 55

Threat researcher Florian Roth recently described the development in more sober terms: until now, human skill, time, and attention limited how many targets an attacker could work. That limit is shifting. As a result, targets that once weren’t worth the effort now pay off for attackers too.

Old weaknesses at machine speed#

Around twenty case reports on ARTEX and related frameworks name the cause of the break-in. Taken together, those causes read like the table of contents of a pentest report:

Table What the break-ins came down to, and what helps

CauseExamples from the casesCountermeasure
Credentials in the wrong place or valid for too longKeys from old leaks, secrets in environment variables, logs and JavaScript files, static tokens for continuous integration and delivery (CI/CD)Secret scanners, rotation, short-lived identities
Forgotten systems exposed to the internetInternal tools and partner portals on the internet, admin endpoints without authenticationMap the attack surface from the outside, put VPN or zero trust in front
Old web and server weaknessesSQL injection, unchecked uploads, XML External Entities (XXE), years-old JBoss vulnerabilityPatch, test applications at the network edge
Too much reach per identityThird-party app that makes itself Global Admin, cloud roles that can run commands on thousands of instancesLeast privilege, cloud audit

In every one of these cases, a capable human hacker could have found and exploited the same holes.

Attacks already worked this way without AI. An analysis (PDF, opens in a new tab) that Volodymyr Styran and Denys Yashchuk wrote for the proceedings of the CyCon 2026 conference shows this. The two draw on just under fifty incident reports from Ukraine’s CERT-UA: in more than half of the cases, the attackers got in without a software exploit, and after initial access the operations ran almost entirely on stolen credentials and built-in tools. In their blog post about it, they write about the AI angle:

“The access-centric reality of cyber operations isn’t being overturned by automation. It’s being industrialised by it.”

Volodymyr Styran and Denys YashchukExploits Were Never the Point, Substack, June 3, 2026

The group VAPOR PANDA shows how such old weaknesses combine. According to CrowdStrike, starting on October 17 it targeted more than 200 organizations in over a dozen countries within eleven days. A break-in has been confirmed at nine of them. At a pharmaceutical company, the attack path ran through web application vulnerabilities to the cloud metadata service and on to a machine identity that was allowed to run commands on thousands of instances. At a Greek government agency, the operator fed the agent a 2022 directory export from the victim’s mail server. The authentication keys in it had probably not been rotated since the export; with them, the agent forged an admin token and read mailboxes. It got into the internal network with VPN configurations from sent emails and a VPN key that reportedly came from a 2024 mail leak.

DIVD, a Dutch volunteer organization of security researchers, is an exception when it comes to initial access. On September 21, the attackers broke in through two previously unknown vulnerabilities in the Zammad ticketing system and had root twelve minutes later. From then on, an ARTEX agent decided on every step itself and, among other things, created an admin token with no expiration date. The vulnerabilities, however, were in an older version of the ticketing system; DIVD admits that it was still running that version. Data leaked out, though the attackers got stuck at the network segmentation. Monitoring fired after 19 minutes, but the team didn’t respond for 19 hours, and another nine hours later the attack was contained, as DIVD writes in its incident report (see Figure 3).

Figure The attackers had root at DIVD after 12 minutes, and the alert went off after 19; then 19 hours passed before the first response.
Speed at DIVD on September 21, 2026, in hours after initial access: the attackers had root after 12 minutes, monitoring alerted after 19 minutes, the team’s first response came after 19 hours, and the attack was contained another 9 hours later. For comparison: for a group tracked by Wiz, Google reports that it usually took about 4 hours from the first account to deep system access.DIVD, case file metricsFor comparison, per GoogleTo rootAlertFirst responseContainedOther group12 minutes19 minutesafter 19 hoursanother 9 hoursusually about 4 hours to system access04 h8 h12 h16 h20 h24 h28 hHours after initial access

After the Hugging Face hack, the victim itself noted in its technical post-incident report that the individual vulnerabilities were familiar. Getting from code execution in a single worker pod to cluster admin across several internal Kubernetes clusters took the agents less than 13 hours. In my secIT digital talk about it on YouTube (in German), I ran a rough estimate: the entire attack chain that the agents worked through in just under four days would have needed a team of three or four specialists.

Why we’re on red alert now#

In my secIT digital talk on YouTube (in German) at the end of September, I still said we were at “yellow alert.” The talk’s title asked whether this was a marketing stunt or a red alert, and since ARTEX and the like, the realistic answer is: we really are on red alert now.

The main reason is speed. It was foreseeable in theory, but it has now shown up in practice again and again. For one attacker group, according to Google Mandiant, it usually took about four hours to get from the first compromised account to deep system access, and in one case 15 minutes. The incident response team at Eye Security watched agents at a customer get around a rate limit in three minutes and come back through Git access that was still open, less than an hour after the environment was back online. In an interview with heise online (in German) in September, based on my own tests with AI attack frameworks, I said that someone with solid basic knowledge who uses an open-source pentest agent goes from initial access to domain admin in your average broken Active Directory in 30 minutes to four hours. In a case from Russia that the security provider RED Security considers very likely to involve ARTEX, the attacker got in through a web application’s default passwords; getting from the first web shell to taking over the domain then took just over two hours. Among other things, the attacker used a password still stored in a Group Policy Object (GPO) and a misconfigured certificate template. On known attack paths, an agent like this sets the pace of a well-drilled team of experienced pentesters.

What it lacks is judgment. In an exposed ARTEX working directory, ThreatMon came across almost 80 findings marked “confirmed,” and none of them had been validated; at DIVD, the agent’s password spraying got in the way of its own man-in-the-middle attack. And Wiz writes that a human attacker of similar skill would hardly spend time testing generic payloads against Microsoft’s heavily scrutinized OAuth flows. Put a skilled hacker at the wheel, though, and with agents they get many times more done than they could before. According to the police in Seoul, a 32-year-old former pentester used an AI coding agent and ARTEX to break into about 60 companies in just six weeks.

On top of that, we still got lucky with PHP2Shell. The two vulnerabilities in PHP leave unpatched web servers open to attack without authentication; the updates came out on October 21.Note 3 Almost two weeks later, according to the Internet of Things search engine Censys, more than half of the 46 million reachable servers running the affected PHP branches still showed a version number without the fix. The vulnerabilities did get exploited, if selectively: on Palo Alto Networks firewalls and Check Point Security Gateways, attackers chained them with other vulnerabilities.

The feared wave of attacks against the millions of ordinary PHP servers, on the other hand, had not materialized by mid-November. Attackers have to use automation to tailor the existing exploits to each target, for example to its memory settings and application. Security researcher and YouTuber LiveOverflow calls it “an easy exploit in 2026” because an agent can adapt it to each target with a single prompt, though he himself points out the difficulty: “It is still a Use-After-Free that requires careful heap grooming and offsets for every target.”

In all likelihood, a lot will go up in flames the moment AI helps uncover a software vulnerability of this caliber that does not depend on the exact version, or once attackers have automated the tailoring to each target further. Pentester Tobias Glemser writes on heise online (in German), translated: “The fact that no working exploit is known for a vulnerability does not mean that it cannot be exploited. For now, it only means that we don’t know of one yet.”

Government agencies had already warned in the summer. The heads of the Five Eyes cybersecurity agencies wrote in June: “The timeline is not years, it is months.” In July, banking supervisors at the European Central Bank (ECB) (PDF, opens in a new tab) required the CEOs of the large banks to submit an action plan against AI-enabled attacks by the end of October. In talks I give to government agencies and operators of critical infrastructure, I see it firsthand these days: anyone who watches the havoc publicly available attack agents wreak in a demo understands, even without another twenty PowerPoint slides with elaborate statistics, that they really have to step on the gas in the coming months when it comes to hardening their own environment.

Stepping on the gas does not mean kicking off a yearlong project with a fifty-page planning document on whether at least your Active Directory might get a green instead of a dark-red PingCastle report by 2028. What it takes is the nerve to run a scream test, with backing from the top (executive board and supervisory board): harden with care, not on the CEO’s computer first and not at month-end, but much faster than before. And if, for all your caution, something breaks, whoever is affected will let you know soon enough.

The chronology of AI attacks now counts more than 560 cases; in about half of them, agents do more than merely assist (see Figure 2). Of its 40 ARTEX cases, we know about roughly one in four only because the attackers slipped up, for example with exposed directories, forgotten consoles, or logs they accidentally sent along.

Where agents hit a wall#

Boring, time-honored security measures stop agents, too. Microsoft counted more than 150,000 failed sign-ins by the group Storm-0940 against just over 300 accounts in eight organizations from October 10 to 14. Account lockout kicked in; not a single attempt succeeded. At the software vendor Dosu, cloud organization policies and hashed API keys kept the damage small. And when agents tried to fool the identity verification at two crypto exchanges with deepfake videos, the verification provider’s server rejected every attempt the agents logged.

Homework, no defensive AI required#

The following list comes from my heise security webinar on offensive AI agents (in German), which I’ll give for the second time on November 30, 2026, from 2:00 to 4:30 p.m. CET. I cross-checked it against the recommendations from Google Mandiant and the FBI (PDF, opens in a new tab). It describes information security basics that work without AI; AI on the defender’s side comes only after that, as the possible (and sensible) next step. Almost every item gets its own post in the blog series; the posts will come out weekly starting in December and will be linked from here.

The list is ordered by importance. If you can only manage two items, take on the first two: in many cases, the agents got in through forgotten systems and credentials left lying around (see Table 1).

  1. Map the attack surface from the outside: What is exposed to the internet, how much of it has to be, and what belongs hidden behind a VPN or zero trust network access, above all the management interfaces of firewalls, hypervisors, and backups? And can anything be opened with default or development credentials?

  2. Find secrets lying around: Comb through repositories, file shares, and container images with secret scanners, rotate everything you find, and take things like API documentation out of production.

  3. Review cloud accounts: Cut back the permissions of roles, service principals, and third-party apps, and replace long-lived keys with short-lived identities.

  4. Patch by urgency: First what is being actively exploited, then what is likely to be exploited.

  5. Harden sign-ins from the outside: Set up account lockout and require MFA for all access from the outside, starting with admins and remote access.

  6. Harden Active Directory and certificate services: Close the attack paths to domain admin and clean up the misconfigured certificate templates that PingCastle and Locksmith uncover, and track down weak and reused passwords; roll out the changes with care but quickly, instead of spending a year planning them.

  7. Detect with built-in tools: Turn on at least minimal logging and consolidate alerts instead of dismissing them one by one; watch for behavior, because a fork may already have removed the tool names.

  8. Deceive: Plant decoys such as honeytokens where agents look first.

  9. Triage alerts in minutes, contain in hours: Have your own on-call team or a managed detection and response (MDR) provider read alerts at night and on weekends, too; decide in advance who is allowed to isolate systems; revoke tokens and check that the revocation takes effect, instead of just blocking IP addresses.

In December, the series starts with setting up the toolbox, a virtual machine with Kali for the tests against your own systems in the following parts. After that comes item 1, the outside view of your own attack surface.

None of this homework is new. This time, though, you really have to sit down and do it.

Further reading#

Read these first:

By me:

For context:

On cloud and secrets:

  • Barracuda: Forty-four hours: the victim itself (“That is on us”), with one of its own keys that a service provider had received and that Barracuda did not decommission after the project, and secrets that ended up encoded in the build log.
  • Datadog Security Labs: ARTEX in AWS accounts, with instructions for searching your own logs.

For detection teams:

  • Unit 42: ARTEX and AI-Powered Red Teaming Tools: which public files help attackers find their way around, plus a table of how to spot agents, such as short bursts of requests followed by a pause of about half a minute for the model.
  • Florian Roth, Beyond Framework Signatures.
  • Andrea Palmieri, part 1 and part 2: uses the ARTEX source code to explain which published indicators do not come from ARTEX at all and which ones a fork has already removed; part 2 does the same for 14 more AI attack tools.
  • Elastic: Inside the loop: five frameworks and their traces, plus an explanation of why revoked credentials do not stop an agent right away.
  • Theori: aipt-detections: rules and queries for Suricata, Sigma, the pattern language YARA, ModSecurity, Microsoft’s Kusto Query Language (KQL), and Splunk’s Search Processing Language (SPL); plus the lab experiment in which ARTEX deleted records without being asked.

Notes

  1. 1The timeline up to the forks draws on StepSecurity, Censys, and Google Threat Intelligence; the figure below links each milestone separately. ↑
  2. 2The counts depend on the day and the method: in early November, Censys counted 1,465 hosts with an exposed ARTEX console (ARTEX: Agentic Asbestos); Check Point found 2,557 systems on which ARTEX was or is running, 1,608 of them publicly reachable (Hacking on Autopilot). ↑
  3. 3The situation and the Censys scan: Kaspersky, PHP2Shell: what we know about the threat two weeks later. ↑