Politico reported on November 19, 2026 that CERT-EU, the cybersecurity service for EU institutions, found the AI pentesting framework ARTEX being used in an attack that has disrupted European Commission staff systems for more than 10 days. What is confirmed is the disruption. What is not confirmed is how the attackers got in, who they are, and whether any staff data was taken. Most of what circulates traces back to one Politico story, which we could not open directly, so this post relies on Cyber Security News's summary and The Next Web's report of it.
| Question | Short answer |
|---|---|
| Am I affected? | Only if you are Commission staff, a contractor on these systems, or run similar HR systems exposed to the internet. Nothing here is a flaw in a coding agent |
| Is there a fix or patch? | None announced. The reported entry point is unknown, so there is nothing to patch yet |
| Was ARTEX the way in? | Unknown. Reporting says it is unclear whether ARTEX helped break in or only searched for weaknesses |
| Was data stolen? | Not confirmed either way |
| Does data leave my machine? | Not because of this story. The risk is to servers ARTEX is pointed at |
What was reported
Per the reporting, a Commission spokesperson, Thomas Regnier, said "individual systems were targeted" and that those concerned had been informed. An EU official told Politico that CERT-EU detected ARTEX's use earlier in November and described it as a publicly available tool that uses AI to look for weaknesses in IT systems. The Corporate Management Board, the Commission's top internal crisis group, reportedly met urgently.
The timeline as reported:
- Early November: CERT-EU detects ARTEX being used. Several applications are taken offline as a precaution.
- Over 10 days: some staff cannot submit timesheets, request leave or enter salary information. Salaries largely arrive as usual.
- November 17: translation services are disrupted.
- November 18: Sysper briefly goes down again; Commissioner Henna Virkkunen briefs fellow commissioners.
- Week of November 19: most affected applications, including Sysper and MyPMO, are back online. The Commission's Digital Services department orders extra security checks.
Claimed vs verified
| Claim | Status | Source |
|---|---|---|
| HR and staff services were disrupted for 10+ days | Reported by multiple outlets, consistent with spokesperson comments | Politico via Cyber Security News, TNW |
| CERT-EU found ARTEX used | Attributed to an EU official, not yet in a public CERT-EU document | Politico via secondary outlets |
| ARTEX caused or enabled the intrusion | Not established; reporting says it may only have scanned | Same |
| Attackers are named | No | None |
| A specific vulnerability was exploited | No | None |
| Staff data was accessed or stolen | Not confirmed | Commission has not confirmed exposure |
| Linked to the January or March incidents | No public link | Separate CERT-EU reporting |
We did not inspect Commission systems or CERT-EU's internal findings. Check the CERT-EU publications page for any formal report, which would supersede this post.
What ARTEX is, and why its presence in an incident matters
ARTEX was published on GitHub by an independent developer in July 2026 under AGPL-3.0. A planner agent turns an operator's goal into attack intents; worker agents claim them and run tools such as nmap, with traffic recorded through mitmproxy. The original repository was removed in late October after abuse reports, but forks, translations and rebrands remain. Our Censys breakdown covers how many consoles sit on the open internet, and the FBI flash on ARTEX covers the government warning.
Detecting ARTEX use is itself a signal. A planner-and-worker loop is noisy compared with a human operator: many tool invocations, systematic enumeration, and fast retries. Unit 42's behavioural hallmarks and Wiz's seven fingerprints describe what that looks like in logs. That may be why CERT-EU spotted it. It also means "ARTEX was used" is not the same as "ARTEX won". A scanner with an LLM planner finds the same exposed services a conventional scanner would, only faster and with less operator skill.
The earlier Commission incidents
The Commission has had a bad year, and it is tempting to string the events together. Public reporting does not. In January, CERT-EU detected an intrusion in the Commission's mobile device management system, with names and phone numbers of some staff possibly accessed. In March the Commission detected an attack on the cloud infrastructure hosting its Europa.eu websites; the Commission and CERT-EU assessed with high confidence that initial access came through the Trivy supply-chain compromise, with a compromised AWS credential used to take files. Those were different systems and different entry points. Treat any claim that all three are one campaign as unverified until CERT-EU says so.
What this means if you run AI coding agents
Nothing in this story says a coding agent was the weak point. The relevance is indirect but practical:
- The tools attackers use are agents too. ARTEX is the same pattern you run on your laptop: an LLM planning, tools executing, a loop repeating. The defenses that matter for your agent (limit what it can reach, log what it ran, review its actions) are the mirror image of what defenders do against an attacker's agent. See what AI agent monitoring tracks.
- Exposed HR and admin portals are the targets. If your team runs internal web apps on the public internet (timesheets, payroll, wikis), assume automated agentic scanning reaches them within hours. Put them behind VPN or SSO, and patch internet-facing software first.
- Credentials sit in agent session files. Operators of attack frameworks have been caught because of leaked session folders; your own
.claudeor Codex session data can hold secrets too. Keep it off web-served paths and out of repos. - Check forks before running them. If you evaluate ARTEX or similar tools for authorized testing, run them in an isolated VM with scoped credentials. Forks of a removed repository can carry anything. For MCP-based tooling, see our MCP security practical guide.
- For defenders: hunt for bursts of sequential requests with varied payloads from one source, a default console on port 8787 on hosts you do not recognize, and unusual API calls to LLM providers from servers.
What Beam can and cannot do here
Beam is local security and observability for coding agents on your own machine. It does not monitor a government's HR servers, and it cannot detect an external ARTEX scan. What it can do is relevant if your developers run agents with access to internal systems: it records what Claude Code and Codex do through hooks, flags credential exposure and destructive commands, redacts secrets before storing events, and can deny risky tool calls via PreToolUse hooks in enforce mode. Its limits: deny is hook-based, not a kernel sandbox; scans of skills and MCP configs are heuristic; and policy fails open if no cached policy exists. See how it works and what it watches.
Why an outage can happen without proven data theft
A 10-day disruption tells you less than it seems to. When defenders find an intruder, or even an aggressive scanner, in systems that hold payroll and leave records, the standard response is to take applications offline, rotate credentials, rebuild from known-good state and run extra checks. That response causes the outage regardless of whether anything was stolen. The Commission's decision to pull several applications offline "as a precaution", as reported, fits that pattern. So does the fact that salaries mostly arrived on time: the payment pipeline was apparently protected while the front-end tools staff use to enter data were not.
This is also why the "claimed" column above matters. Outage length, an AI tool in the headline and an HR system together make a dramatic story, but each of the three facts could have an ordinary explanation. A long recovery can reflect careful rebuilds. An AI tool name can reflect a scan that found nothing. An HR system is a common target because it is a web application with many users, not because anyone wanted payroll specifically.
Questions people are searching
Is it "AI hacking" the Commission? Reporting describes a human-run attack in which an AI-driven tool was observed. No source says the AI acted without an operator.
Is the Commission's data safe? The Commission has said individual systems were targeted and those concerned were informed. It has not published findings on exposure. If you are affected, expect direct notification and watch for phishing that references payslips or leave requests, which criminals often send after an HR-system incident.
What should an organisation do now? Inventory internet-facing admin and HR portals, confirm patch levels, require multi-factor sign-in, and make sure logs from those systems go somewhere an incident responder can search. Then check whether your detection can spot a burst of systematic scanning, since that is the one artefact reporting says gave the tool away.
What to watch next
Three things would settle the open questions: a CERT-EU or Commission technical note naming the entry point, a statement on whether staff data was accessed, and any attribution. Until then, the defensible summary is "an AI pentest tool was observed during an incident that disrupted services", not "AI hacked the Commission". For the wider pattern of November incidents, see our November 2026 attacks timeline.
Frequently asked questions
Did hackers use ARTEX against the European Commission?
According to Politico, as relayed by Cyber Security News and Reuters, CERT-EU detected ARTEX being used earlier in November. It is not established whether ARTEX was used to break in or only to scan for weaknesses, and the Commission has not published a technical report.
What is ARTEX?
ARTEX is an open-source, LLM-driven penetration testing framework published on GitHub in July 2026 under AGPL-3.0. A planner agent breaks a goal into attack intents and worker agents run tools such as nmap. The original repository was taken down, but forks and rebrands circulate.
Is European Commission staff data stolen?
Not confirmed. Reporting says investigators have not established whether anyone accessed or stole staff data, and the Commission has not confirmed that payslips, leave records or medical reimbursement data were exposed.
Which Commission systems were affected?
Reporting names Sysper (payslips, working time, leave) and MyPMO (medical reimbursements), plus a translation-service disruption on November 17. Most affected applications are reported back online.
Is this the same as the earlier Commission breaches this year?
There is no public link. A January attack hit mobile device management and a March cloud breach was tied by CERT-EU to the Trivy supply-chain compromise. Neither has been connected to this incident.
Is ARTEX safe to run on my own machine?
It is a dual-use offensive tool that drives other tools autonomously. Run it only against systems you own or are authorized to test, and in an isolated environment. Treat forks of unknown origin as untrusted code.
Related reading
- ARTEX on the open internet: Censys findings
- FBI flash on ARTEX
- Unit 42 on ARTEX behavioural hallmarks
- Censys report on ARTEX
Details here come from secondary reporting of a Politico story and may change as the Commission and CERT-EU publish more. Accurate as of November 19, 2026. Beam is a local prototype; its documented limits are stated above.

