European Commission Cyberattack: Ten Days Offline and the ARTEX AI Pentesting Framework Behind It

The CyberSec Guru

European Commission Cyberattack

If you like this post, then please share it:

Buy me A Coffee!

Support The CyberSec Guru’s Mission

🔐 Fuel the cybersecurity crusade by buying me a coffee! Why your support matters: Zero paywalls: Keep the main content 100% free for learners worldwide.

“Your coffee keeps the servers running and the knowledge flowing in our fight against cybercrime.”☕ Support My Work

Buy Me a Coffee Button

For more than ten days in November, the European Commission could not reliably do what an employer has to do. Sysper, the internal platform where staff check payslips, file leave requests and record working hours, was taken offline as a precaution. MyPMO, the Paymaster Office portal that processes medical reimbursements under the Joint Sickness Insurance Scheme, went down with it. Translation services failed on Tuesday. Both HR systems came back earlier in the week, and Sysper crashed again on Wednesday.

Politico’s Brussels Playbook first reported the disruption on 19 November, and Techzine summarised it. It reached the top of the institution. The Corporate Management Board, the Commission’s senior internal crisis body, held an urgent meeting on “ongoing cyber incidents” earlier in the month. The digital services department ordered stepped-up security checks. Executive Vice-President Henna Virkkunen, who holds the tech sovereignty, security and democracy portfolio, briefed the College of Commissioners on Wednesday.

The outage is the least interesting part of the European Commission cyberattack. The tooling behind it is the reason to read on. According to Politico’s sources, CERT-EU, the cybersecurity service for the EU’s institutions, detected earlier this month that attackers had used ARTEX, a freely available AI-driven penetration-testing framework, to scan the Commission’s systems for weaknesses. “Some actors misuse such tools and have targeted the Commission,” an EU official told the newsletter. Spokesperson Thomas Regnier confirmed that “individual systems were targeted” and that, in line with standard practice, those concerned had been informed.

Whether staff personal data was exfiltrated is still open. The Commission believes it has contained most of the attack, but the full extent of the damage is unclear. That uncertainty is what a forensic process honestly produces when it has to separate reconnaissance from exploitation from theft across systems that were deliberately powered down mid-investigation.

What the European Commission cyberattack confirms, and what it does not

The story is being told loosely in places, so precision helps. CERT-EU has not named an attacker or disclosed a specific vulnerability. It has not said publicly whether ARTEX enabled the breach itself or was used mainly for reconnaissance. Its advisory of 5 November said only that it had observed AI-driven activity against its constituency that “highly likely involved the use of ARTEX,” alongside campaigns by several threat actors worldwide. The Commission’s own statements stop at “a detected cyber campaign.”

What can be said with confidence is that the same framework has been documented in forensic detail against banks, cloud tenants and hundreds of other organisations. Its mechanics explain the shape of the Commission’s disruption and why the investigation is hard.

What is ARTEX? The AI pentesting framework explained

ARTEX is not a scanner with a chatbot bolted on. It orchestrates the part of an intrusion that used to need a skilled human: deciding what to try next. Censys’s analysis describes a Go binary with a Next.js web console backed by PostgreSQL. An operator defines a Task with a scope, test accounts and a natural-language goal. A Planner agent breaks that goal into “intents”. Worker agents claim the intents, run them with ordinary offensive tooling (the Docker image ships with nmap) and write results back. An event-driven engine repeats the explore, decide, execute loop until the goal is met or the routes run out.

Meanwhile the system keeps two graphs, an asset graph of what the target looks like and an exploration graph of how much of it has been tested. It records traffic through mitmproxy and exposes a dashboard for task orchestration, findings, coverage and an approval queue, so a human can watch and steer. Qualys, which read the source, gives the simplest accurate description: a project manager for a break-in.

Where ARTEX came from and why deleting it did not help

The author published it on GitHub as Autumn-27/ARTEX on 26 July 2026 under AGPL-3.0. The same developer wrote ScopeSentry, a widely used Chinese asset-discovery platform, and ARTEX syncs ScopeSentry’s domains, subdomains, IPs, ports and endpoints straight into its asset graph. It reached an estimated 1,500 to 2,000 stars within ten weeks.

It ships no model of its own, so the operator connects an external LLM. The FBI’s FLASH advisory lists integrations with Anthropic Claude, OpenAI GPT, DeepSeek, GLM, Grok and Kimi. It also notes that attacker-controlled instances overwhelmingly favour DeepSeek’s Flash-family models, for their cost and because no guardrails are observable when the work is framed as penetration testing.

📬 Stay Ahead of Cyber Threats

Get the latest cybersecurity news, critical vulnerabilities, threat intelligence, tutorials, and exclusive giveaways delivered straight to your inbox. No spam. Unsubscribe anytime.

CrowdStrike published evidence of malicious use on 7 October, and the author removed the repository on 8 October, saying the tool was meant for authorised testing. Removal did not remove the tool. It survives in forks, archived copies, localised rebrands such as ScopeWeaver and a commercial edition, ARTEX PRO, which the FBI notes keeps it available to anyone who wants it.

ARTEX has no hard scope enforcement by design

Defenders should look hard at the governance design. CERT-EU states that ARTEX relies on configurable rules and the model itself to respect scope and avoid destructive actions, with no hard technical enforcement. Qualys goes further. The framework did not ship without an authorization-scope mechanism; it shipped with one removed, and the comment saying so is still in the source.

The gates that replaced it are seeded as ordinary settings a user can disable or delete, and when a safety check errors out or times out, the default is to carry on. The pause between one step landing and the next being assigned is a constant of 800 milliseconds. The planner is told to put depth ahead of coverage, pushing a promising route all the way through instead of testing everything lightly. Capability and permission are decoupled by design. What separates an authorised red team from an intrusion is operator intent and a settings file, not architecture.

What ARTEX has already done: documented attacks

The public evidence is more measured than the headlines, and more useful.

CrowdStrike: ARTEX against South Korean financial organisations

CrowdStrike’s 7 October report documented a campaign against South Korean financial organisations from late September to early October. The actor breached a broker-facing loan inquiry service at one bank and an employee mobile work-support system at another. It ran ARTEX on 38.244.50[.]120 with DeepSeek v4.1-flash as the primary backend, supplemented by GLM-5.3 and Grok 4.6 through Claude Code sessions, and likely bought model access through the reseller xcai[.]pro.

CrowdStrike found the operation because the actor left open directories exposing a .claude folder, Chinese-language pentesting prompts, session histories and ARTEX configuration files. The FBI puts the personal information exfiltrated in the South Korean financial attacks at approximately 67,000 individuals.

Zscaler: inside the operator’s AI-assisted workflow

Zscaler ThreatLabz reconstructed the operator’s working environment from prompts spanning 4 August to 4 October, and it reads like a mature, if criminal, engineering shop. A work server held project files and a local ARTEX deployment, and a jump server routed connections to targets. Custom rsync scripts synchronised instructions and memory files across accounts configured for different models, with GLM assigned to review DeepSeek’s work, and its subagents’ work, for gaps.

Each investigation lived in a folder with scope.md, surface.md, notes.md, findings and evidence. A hook script injected project context into every new prompt so sessions could resume where others left off. Instructions told the models to assume all supplied targets were preauthorized and to skip per-step approval. Targets spanned roughly nineteen countries and sectors from dating apps to SMS providers to legal settlement systems, with an explicit focus on structured personal data and phone numbers. Prompts asked where Korean breach data is usually sold, and a standing warning said not to initiate direct withdrawals because those trigger alerts.

Microsoft: Storm-0940, ConnectWise and ARTEX

Microsoft attributed the ConnectWise Cloud Backup supply-chain compromise to Storm-0940, a China-state-aligned actor active since 2021. The actor used ARTEX as early as September to automate scanning, password spraying and OAuth application abuse against hundreds of organisations.

Between 10 and 14 October, infrastructure at 144.172.64[.]99 sprayed 327 accounts across eight organisations, generating more than 151,000 failed logons and peaking at 2,900 failures per minute. Forty-eight accounts absorbed over a thousand failures each, and lockout responses made up nearly 98 percent of activity. None of it succeeded.

After ConnectWise’s 16 October advisory, the same actor used stolen credentials for already-trusted multi-tenant applications to authenticate app-only into customer environments. It attempted sign-ins at more than 2,250 tenants and, in several cases, went from first targeting to privileged cloud access in under three hours. Follow-on activity through Microsoft Graph included reconnaissance, credential additions to OAuth applications, privilege changes and bulk secret reads across Azure Key Vaults. Microsoft warns that those credential modifications can persist as a backdoor even after the original compromised application is remediated. It also stresses that an affected application does not by itself prove data was accessed.

How many ARTEX consoles are exposed on the internet?

Censys counted 1,465 hosts exposing an ARTEX console in early November. About 650 sat in two identically configured hosting blocks, which leaves roughly 815 distinct-looking deployments. The FBI, citing researchers, puts the exposed-console figure above 1,600, and both note that the real population behind firewalls is larger.

Some 155 hosts, 10.6 percent, also ran other detected threat tooling, including CyberStrikeAI, PentAGI, Cobalt Strike, Vshell and AdaptixC2. Forty-three ran self-hosted LLM gateway or reseller software such as CLIProxyAPI and New API. Thirty-two open directories on ARTEX hosts exposed post-exploitation kits, stagers and implants. Censys scan history shows the CrowdStrike-named host had been leaking its operator’s files since 18 August, seven weeks before anyone published.

Qualys adds that CrowdStrike observed one group point the framework at more than 200 organisations in eleven days, with confirmed follow-on activity at nine and a human steering at decision points. Microsoft separately timed an agent at five minutes from searching GitHub for leaked secrets to pulling data from a cloud key vault.

Three hours against seven days: why patch cycles lose to agentic attacks

The strategic content is arithmetic. Qualys reports that its customers now fix critical vulnerabilities in a median of seven days, down from nine a year earlier, with 12 to 16 percent remediated before CISA adds the flaw to its Known Exploited Vulnerabilities catalog. That is real progress, and it is still 168 hours against an attacker who needs three.

Most vulnerability programmes rest on an assumption that rarely gets written down: attacker time is scarce. It is why a critical bug on an internet-facing server gets thirty days, why a local privilege escalation waits behind the remote ones, and why risk exceptions get signed when exploitation looks complicated. ARTEX makes attacker time cheap, and the assumption goes with it. CERT-EU makes the same point from the tester’s side. The framework can check large numbers of low-complexity misconfigurations that a human would notice but not have time to verify, work through long attack chains without fatigue, and craft the tailored payloads that manual assessments skip because the effort exceeds the severity of the finding.

Two consequences follow for prioritisation. First, the planner works depth-first along routes, so two medium findings on the same path are worth more to it than a critical on a server that leads nowhere. A queue sorted by severity alone keeps fixing the wrong thing first. CERT-EU’s advice to retain non-severe findings in the remediation cycle exists to stop minor misconfigurations being chained into high-impact paths. Second, each step depends on what the previous one produced, so removing one link stops everything behind it on that route. That is why exposure reduction and credential hygiene buy more than their CVSS scores suggest.

The FBI’s observed technique list is a catalogue of the unglamorous: unsigned SAML assertions, XXE, SSRF, IDOR, SQL injection, command injection, hardcoded administrator passwords in SPA JavaScript bundles, missing authentication on web APIs, stolen VPN configurations, web shells deployed via MySQL log poisoning, Kerberoasting and DCSync, LAPS password reads, enumeration of more than 110,000 Active Directory objects, and even inventorying endpoint sensor coverage to find Windows machines without CrowdStrike Falcon. None of it is new. All of it is now tried at once, around the clock, by operators who may not understand it.

The backdoored ARTEX Docker image: tomicdd/artex

After the author pulled the repository on 8 October, users went looking for replacements. They found the most-starred copy, whose source matched upstream and whose installer pointed at a Docker image called tomicdd/artex, with a byte-identical re-push at congdonglinux/artex.

BadByte’s analysis, published 4 November and detailed by Ostorlab, found the image carried one added function, a 160-byte verifyCaptcha, called from the login handler right after the JSON body is decoded and before any credential check. A login request whose captcha field begins with a secret marker is passed to sh -c as the container’s root user, and the command output comes back in the JSON response. No account, no password, no session. The container runs as root and holds the PostgreSQL superuser password, LLM provider keys, notification credentials, engagement targets, harvested credentials, findings, full intercepted traffic, the mitmproxy CA private key and the JWT signing key that forges seven-day admin tokens.

The timeline is tight enough to suggest intent. The Docker account was created at 18:50 UTC on 8 October and the image pushed at 19:41, hours after the original repository disappeared at 10:50. The first backdoored console came online on 20 October. A GitHub issue titled “suspect a backdoor” was opened on 23 October and closed the next day. BadByte counted 62 internet-facing consoles serving the build between 2 and 4 November, 7,946 pulls of the two images by 3 November, 9,077 by 5 November, and 637 repositories still pointing at them. There is no CVE and no KEV entry, because no active maintainer exists to issue an advisory, and the tool’s own update path pulls the same poisoned image again.

Some of this backdoor’s victims are almost certainly defenders and researchers evaluating the tool. That is why Qualys now tells organisations to inventory AI pentest frameworks in their own estates and their suppliers’ estates as software assets, approve or remove them, and rotate every secret those tools could read. The offensive ecosystem has no clean supply chain either. It is a commons with predators in it.

The Commission’s three 2026 security incidents: Ivanti, Trivy and ARTEX

The November intrusion is the Commission’s third publicly disclosed security event of the year, and together the three map the modern attack surface better than any one does.

On 30 January, its central mobile-device management infrastructure identified traces of a cyber-attack that may have exposed staff names and mobile numbers. The Commission said the system was contained and cleaned within nine hours and no mobile devices were compromised. Reporting tied the incident to exploitation of Ivanti Endpoint Manager Mobile vulnerabilities, the classic edge-device initial-access pattern.

March brought a different class. CERT-EU’s write-up describes how a compromised Trivy release, the supply-chain attack tracked as CVE-2026-33634 and publicly attributed by Aqua Security to TeamPCP, handed an actor an AWS API key on 19 March. The actor launched TruffleHog to hunt for more secrets, attached a new access key to an existing user to evade detection, and exfiltrated data from the backend of the europa.eu hosting service. ShinyHunters published the dataset on 28 March: roughly 91.7 GB compressed, 340 GB uncompressed, touching up to 71 clients including at least 29 other Union entities. The Commission notified the European Data Protection Supervisor, and its press release stressed that internal systems were not affected.

November reverses that reassurance. This time the affected systems are the internal administrative estate itself, the target class the FBI and Akamai describe ARTEX as excelling against: partner and supplier portals, mobile backends, employee applications, sales-support systems, legacy services and third-party integrations, the peripheral systems that get less security attention than flagship applications.

Sysper holds payroll, leave and timesheet data. MyPMO holds medical reimbursement claims, which are special-category health data under the EU’s data-protection rules for its own institutions. An institution with a broad, heterogeneous digital estate deeply connected to third-party services is where an agent that never gets bored and never skips the dull target earns its highest expected return. March showed the Commission’s trust in third-party tooling could be weaponised. November shows its own forgotten corners being worked over at machine speed.

EU breach notification rules: Regulation 2018/1725, the EDPS and the 72-hour clock

Much commentary has framed the Commission’s cautious communication as a GDPR problem. The law is more specific. As an EU institution, the Commission operates under Regulation (EU) 2018/1725 rather than the GDPR, and the European Data Protection Supervisor is its regulator. Article 34 requires notification of a personal-data breach to the EDPS without undue delay and, where feasible, within 72 hours of becoming aware of it, unless the breach is unlikely to result in a risk to individuals. Article 35 requires communication to data subjects without undue delay when the breach is likely to result in a high risk.

Separately, Regulation (EU, Euratom) 2023/2841 governs incident reporting to CERT-EU: an early warning within 24 hours of becoming aware of a significant incident, an incident notification within 72 hours, and a final report within a month of that notification, with user communication obligations where affected users need to take mitigating measures. An incident is “significant” if it causes or can cause severe operational disruption, which ten days without HR systems plainly is.

These clocks start at awareness, and awareness is a forensic judgement, not an alert. When the tooling is designed to move laterally, escalate and exfiltrate with minimal human interaction, the trail fragments across systems. The difference between attempted access and successful theft may stay unresolved until databases and logs are worked through. Microsoft’s guidance after ConnectWise makes the same operational point: successful authentication by a compromised application does not necessarily mean customer data was accessed or exfiltrated.

The Commission could not yet say whether personal data was compromised, so it could not say what had to be notified, to whom, or under which article. That is legally coherent. It is also why staff experienced the incident as silence. Techzine reports that a Sysper alert referred only to a general problem without mentioning a cyberattack, and that several officials complained about the communication. Notification regimes built around a 72-hour awareness threshold now sit upstream of investigations into attackers who work in three-hour cycles, and the gap lands on the people whose data it is.

How to defend against ARTEX: CERT-EU and FBI guidance

The guidance that has accumulated around ARTEX is unusually specific, and it converges on one theme: shrink the time and the surface an agent can work with.

CERT-EU’s immediate measures are to inventory internet-exposed assets and limit exposure to what genuinely requires it, isolate management and control panels from untrusted networks, prioritise emergency patching of everything internet-facing, and watch privileged and administrator accounts closely. Its short-term measures add MFA across critical systems, hunting for connections to known actor infrastructure and for ARTEX-consistent scanning patterns and tool signatures, and EDR coverage of unusual process behaviour.

The FBI adds application-layer controls against automated enumeration: rate limits, recipient limits and transaction thresholds on high-impact business APIs; alerts on sequential lookups and unusually high record counts per session on partner and employee portals; server-side authorization on customer records; and sanitised error responses that stop leaking stack traces, SQL errors and framework details to unauthenticated callers.

How to detect ARTEX activity: LLM API calls, gateways and agent signatures

Several detection opportunities are peculiar to agentic tooling. The FBI lists outbound calls to LLM API endpoints such as api.anthropic.com, api.deepseek.com, open.bigmodel.cn and api.x.ai as worth investigating alongside other suspicious activity. It also lists self-hosted LLM gateways or reseller software in egress logs, token-consumption and API-billing anomalies, high-frequency sequential enumeration against login and search endpoints, structured personal data moving outbound over HTTPS to unauthorised destinations, and .aspx or .jsp files in unexpected web paths.

It even asks reporters for evidence of multiple sub-agents overwriting each other or generating duplicate requests, a forensic signature of parallel agents no human operator produces. Censys’s hunting guidance pivots on the default console port 8787, co-located open directories, Vshell or FRP listeners and LLM proxy software on the same host. None of this needs new technology. It needs logging the estate you actually have, including the AI tools your own staff and suppliers run.

Faster patching and containment against AI-driven intrusion

The harder change is tempo. Qualys argues that vulnerability programmes must detect exposures at disclosure, rank them by confirmed exploitability instead of severity alone, make automatic patching the default with deadlines of hours for internet-facing systems, and reopen every risk exception justified by attacker effort, because the impact column has not moved but the likelihood column has. Its telemetry says the gap is closable: customers deployed more than 180 million patches in twelve months with fewer than one in 8,000 needing rollback, and the share running fully automated roughly doubled over the year.

Containment needs the same treatment. Taking Sysper and MyPMO offline was a blunt but reliable manual primitive, and almost certainly the right call under uncertainty. Whether it reflected absent automation or a deliberate choice is not public. What is public is that recovery is a process, not a state, and Sysper’s second crash on Wednesday is a reminder that bringing systems back under pressure, with residual-compromise risk unresolved, deserves the same rigour as the shutdown.

Outages versus attacks: the Optimizely supply-chain compromise

The Commission’s incident landed in a week that confused failure with attack across the Western technology estate, and the distinction is worth keeping. The outages reported by CNBC and others were mostly not attacks. Microsoft’s 12 to 14 November disruption was a latent software defect in networking gateway servers. The FAA’s 13 November ground stop traced to telecommunications instability affecting legacy equipment. JPMorgan Chase’s 16 November transaction delays were a software bug the bank said was not cyber-related. The NSA’s internal communications outage on 17 November had no reported evidence of a cyberattack.

The genuine attack that week was a supply-chain compromise of another kind. Scott Helme of Report URI found that on 16 November, malicious JavaScript was served through Optimizely’s experimentation snippets to visitors of 178 organisations, including the BBC, CNN, Zoom, Air France-KLM, HSBC and Ledger. It swapped cryptocurrency addresses in pages and clipboards, skimmed card data, stole cookies and page screenshots, and loaded iOS exploit kits, all from a CDN everyone trusted. It was cleaned within hours. How the attacker gained write access to so many customer projects remains unexplained.

What the Commission breach shows about AI-driven attacks

Offensive capability is being commoditised into free, forkable, sometimes booby-trapped software that automates the skilled connective tissue of an intrusion. The institutions defending against it still schedule scans by the week and patch by the change window.

The Commission is both an operator of sensitive administrative systems and the regulator shaping the EU’s response, having introduced a Cybersecurity Package in January built around a Trusted ICT Supply Chain framework, NIS2 and the Cyber Solidarity Act. The ARTEX episode tests whether that emphasis on notification and accountability is matched by investment in exposure management, privilege minimisation and containment that can move at the speed of what it contains. The Commission has up to a month to file its final incident report. The framework that reportedly probed its systems needs about three hours to reach privileged access.

Buy me A Coffee!

Support The CyberSec Guru’s Mission

🔐 Fuel the cybersecurity crusade by buying me a coffee! Your contribution powers free tutorials, hands-on labs, and security resources.

Why your support matters:
  • Writeup Access: Get complete writeup access within 12 hours
  • Zero paywalls: Keep the main content 100% free for learners worldwide

Perks for one-time supporters:
☕️ $5: Shoutout in Buy Me a Coffee
🛡️ $8: Fast-track Access to Live Webinars
💻 $10: Vote on future tutorial topics + exclusive AMA access

“Your coffee keeps the servers running and the knowledge flowing in our fight against cybercrime.”☕ Support My Work

Buy Me a Coffee Button

If you like this post, then please share it:

News

Discover more from The CyberSec Guru

Subscribe to get the latest posts sent to your email!

Leave a Comment

This site uses Akismet to reduce spam. Learn how your comment data is processed.