Ai Agents 6 min read

European Commission confirms an 'AI-generated' cyber campaign; reports tie it to ARTEX after 10 days of HR outages

The Commission said on November 19 that it took staff applications offline after detecting an AI-driven intrusion campaign on October 22. Politico's sources say CERT-EU found the attackers using ARTEX. What is confirmed, what is not, and what the FBI and CERT-EU advisories say defenders should do.

The European Commission confirmed on Thursday, November 19, that it had responded to what its spokesperson called an “AI-generated cyber campaign” against its internal IT systems. The confirmation followed a report that morning in Politico’s Brussels Playbook, which said the Commission’s staff HR and medical reimbursement platforms had been disrupted for more than 10 days and that defenders had detected the attackers using ARTEX, the open-source AI penetration-testing framework behind a run of intrusions this autumn.

It is rare for a government institution to describe an incident against itself in those terms. It is also a case where the on-record facts and the reported ones differ in ways that matter, so they are separated below.

What the Commission has confirmed

Spokesperson Thomas Regnier gave the on-record account, reported by The Record and The Register:

  • The Commission detected the first signs of the campaign on October 22.
  • It “temporarily took down several applications” as a protective measure, and most have since been restored. The outages were the defenders’ doing; no report so far says the attackers knocked systems over.
  • “Individual systems were targeted,” and the people concerned have been informed. There is no evidence of a data breach affecting staff broadly, and no ransom demand.
  • Other systems were checked preventively, and the Commission is working with CERT-EU (the EU institutions’ incident response team) and has contacted the European Data Protection Supervisor.
  • An investigation is open. The Commission has not named the tool, the attackers or a motive.

Asked why none of this was disclosed for four weeks, Regnier said publicizing defensive measures could have undermined them, adding: “There was no hiding whatsoever.”

What is reported but not confirmed

The rest comes from Politico’s sources and other outlets’ follow-ups:

  • The systems. Sysper, which staff use for payslips, timesheets and leave, and MyPMO, which handles medical reimbursements, were offline for more than 10 days. Both came back early this week; Sysper went down again briefly on Wednesday, November 18, and translation services were disrupted on Tuesday. Salaries were paid as normal.
  • The tool. Politico’s sources say CERT-EU detected the attackers using ARTEX in early November. French outlet IT-Connect notes it is not known whether ARTEX was used to break in or only to hunt for weaknesses once the attackers were probing.
  • The data. The affected platforms hold officials’ names, pay grades, home addresses and contact details, according to EU officials cited by Euronews. Whether attackers read or copied any of it, including payslips, has not been established. “No evidence of a broad breach” and “some individuals were notified” are both true at once, which suggests a limited confirmed exposure and an unfinished investigation.
  • The escalation. The Commission’s Corporate Management Board held an emergency meeting earlier in November, and tech commissioner Henna Virkkunen briefed fellow commissioners on Wednesday.

No source gives an initial access vector, a CVE, or any attribution.

ARTEX, and the warnings that preceded this

ARTEX takes an objective in natural language, splits it across agent workers that run real reconnaissance and exploitation tools, and reports to an operator dashboard. It ships without a model: operators plug in an LLM provider of their choice. We covered its mechanics and its role in the Chainstack and AWS customer breaches in our earlier ARTEX explainer. The original author pulled the GitHub repository in October after misuse reports; what circulates now is forks.

Two government advisories landed before the Commission went public:

  • CERT-EU, November 5. Its post, “From Penetration Testing to Cyber Threat: Beware of ARTEX”, said it had “observed similar instances of AI-driven activities against its constituency” that “highly likely involved the use of ARTEX.” CERT-EU’s constituency is the EU institutions, so with hindsight that sentence was written two weeks into the Commission incident. The post also makes a point worth keeping: ARTEX’s scope rules, the feature meant to keep a pentest inside authorized targets, are not technically enforced.
  • FBI, November 13. An FBI advisory reported more than 1,600 ARTEX consoles exposed to the internet and intrusions that went from target identification to privileged access in under three hours, as summarized by The Register and The Record.

CERT-EU maps the activity to three ATT&CK techniques: active scanning (T1595), obtaining a tool (T1588.002) and exploiting public-facing applications (T1190). None of that is novel tradecraft. What changes is the cost: an operator no longer needs the skill to chain reconnaissance, exploit selection and privilege escalation by hand. Horizon3.ai CEO Snehal Antani told The Register that chaining which took a skilled operator days or weeks “is now happening in hours,” and that shared tooling will make attribution harder because state-backed groups and criminals end up using the same framework.

Why a four-week timeline matters

The dates describe a defender that was slower than the tool:

DateEvent
October 22Commission detects first signs of the campaign
Early NovemberCERT-EU detects ARTEX use, per Politico’s sources; management board meets
November 5CERT-EU publishes its ARTEX warning without naming the Commission
November 13FBI advisory on ARTEX
November 17-18Translation services disrupted; Sysper drops again after restoration
November 19Politico reports the incident; Commission confirms

If the FBI’s three-hour figure is typical, the gap between first detection and a contained environment here was hundreds of times longer than the attack loop. Taking HR and reimbursement systems offline for 10 days is what containment looks like when you cannot quickly tell which hosts an automated operator has already touched. The Sysper relapse a day before disclosure suggests that work was not finished.

It is also the Commission’s third disclosed security incident of 2026, after zero-day exploitation of its mobile device management infrastructure in January and a cloud hosting breach in March, per The Record. Officials have not linked them.

What defenders should take from it

CERT-EU’s recommendations are ordinary, which is the point: an agentic framework mostly finds what a patient human would find, sooner.

  • Shrink and inventory the internet-facing surface, and isolate management panels. T1190 is the entry technique CERT-EU lists.
  • Treat emergency patching windows as hours. The GTIG report on patch-diffing agents showed the same compression on the exploit-development side.
  • Sanitize error responses. Verbose errors are free context for an LLM planner in a way they never were for a scanner.
  • Hunt for ARTEX artefacts and exposed consoles, including your own red team’s. The FBI’s 1,600 exposed dashboards are other people’s footholds waiting to be borrowed.
  • Monitor privileged accounts and enforce MFA, since fast privilege escalation is the measured outcome.

What to watch next: whether the Commission or the data protection supervisor confirms what staff data was accessed, whether CERT-EU publishes indicators from the incident itself, and whether anyone names the model provider the operators used. That last detail decides whether this becomes a question for an AI lab’s abuse team as well as for Brussels.

Get Insanely Good at AI

Get Insanely Good at AI

The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.

Keep Reading

Ai Agents

How to Secure AI Agents With Google ADK

Learn how to secure your autonomous workflows and prevent unauthorized actions using Google ADK's hardware-backed tool binding and execution logs.

Ai Agents

Cloudflare: 50+ AI pentest frameworks scanned its customers in three days, and traffic is up 30x since October 1

Cloudflare's November 13 report says taking ARTEX offline produced a hydra effect: hundreds of forks, 50+ framework families and a 30x rise in agentic attack traffic. Here is what the data shows, what it leaves out, and what a stolen CI/CD token looked like in one incident.

Ai Agents

ARTEX in the wild: Chainstack's breach and the AntiFox fork show what agentic attack frameworks actually do

Chainstack says an attacker using the ARTEX agent framework went from a default-password Grafana to production secrets in five days, and Arctic Wolf ties an ARTEX fork called AntiFox to AWS secret theft. Here is how the intrusions worked and what defenders should look for.

Ai Agents

Perplexity's sandbox tests: zero VM escapes from 12 models, and the cyber-tuned ones did worse

Perplexity's November 11 follow-up pits 12 frontier models against its own agent sandbox. None escaped the Firecracker VM, five beat the old network policy every time, and the cyber-focused GPT-5.6 Cyber and Claude Mythos 5.1 trailed general-purpose models.

Ai Agents

SSHole: agent swarms found a ten-year-old OpenSSH login bypass, and attackers were using it within two days

CVE-2026-123158 lets a 257-byte password log in to accounts that have no password on musl-based Linux such as Alpine. A researcher found it by having AI agents fire about 17,600 odd logins at differently built sshd binaries. OpenSSH 10.6p2 fixes it.