⌁AI·CYBER·BRIEF▌
AI Threats11 min read

The ARTEX takedown backfired: Cloudflare sees 30x more AI attack traffic, and the EU Commission was hit

Pulling ARTEX offline spawned hundreds of forks. Cloudflare and FBI data, plus the EU Commission breach, show why defenders must treat AI attack agents as permanent.

Taking the ARTEX repository offline on October 8 did not slow AI-driven attacks. Cloudflare says the opposite happened: traffic from AI penetration-testing orchestrators has grown more than thirtyfold since October 1, and ARTEX plus its forks now make up over 60% of the orchestrators it knows about. The European Commission's confirmation on November 19 that it was hit by an "AI-generated cyber campaign" shows that targets now include large, well-defended institutions, so defenders should plan for these tools as normal background traffic, not wait for a takedown to remove them.

The short version

ARTEX is a free, open-source "autonomous penetration testing" tool. Penetration testing means attacking your own systems to find weaknesses before criminals do. ARTEX hands that work to AI agents: the operator picks a target, and a team of agents running on commercial AI models maps the systems, tries known weaknesses, steals credentials and moves deeper. It contains no AI model of its own and rents its "brain" from whichever AI provider it is configured to use.

Its developer, who posts as "Autumn-27", pulled the original GitHub repository on October 8, 2026. The usual expectation is that removing a popular tool slows its spread. On November 13, Cloudflare's threat-intelligence team Cloudforce One published data pointing the other way. The takedown "likely caused a hydra effect": copies, forks and look-alike frameworks multiplied, and the attack traffic Cloudflare tracks went up, not down.

On November 19, the European Commission confirmed it had been hit by a cyber campaign first detected on October 22. Staff systems for payroll, leave and medical reimbursement were disrupted for more than ten days. Reports say the EU's own cyber-response team, CERT-EU, linked the attack to ARTEX. The Commission itself has not named the tool.

For anyone in IT, the practical point is simple. These agents are not inventing new attacks. They are very fast, very patient and very cheap at finding the doors that are already open: an exposed admin panel, a key that never expires, a forgotten test server.

Timeline

  • July 2026 — ARTEX is first published on GitHub by "Autumn-27", according to the FBI.
  • August 2026 — The FBI says malicious ARTEX use has been observed since at least this month.
  • October 1, 2026 — The start of the period over which Cloudflare measures a more-than-30x rise in daily requests from AI pentesting tools.
  • October 8, 2026 — The developer removes the original ARTEX repository. Forks, mirrors and a commercial "ARTEX PRO" keep circulating (FBI).
  • October 22, 2026 — First signs of the campaign against the European Commission are detected, spokesperson Thomas Regnier said, as reported by The Record.
  • Early November 2026 — CERT-EU detects ARTEX use in the Commission attack, according to a European official quoted by Politico (via IT-Connect). The FBI says more than 1,600 ARTEX consoles were publicly exposed at this point.
  • November 12, 2026 — Cloudflare's cut-off date for its statement that ARTEX and forks make up over 60% of known AI pentest orchestrators.
  • November 13, 2026 — Cloudflare publishes its analysis. The FBI publishes FLASH-20261113-001 on ARTEX, the same day as Amazon's report on ARTEX agents stealing AI keys.
  • November 17–18, 2026 — Commission translation services are disrupted on November 17, and the Sysper HR system briefly fails again on November 18 (IT-Connect).
  • November 19, 2026 — Politico's Brussels Playbook reports the Commission attack. The Commission confirms it the same day.

What we know

Cloudflare's numbers. The post is by Blake Darché, Myles Sloan and Jacob Crisp of Cloudforce One. Its key claims, in Cloudflare's words:

  • "Daily requests of AI penetration testing (pentesting) and exploitation tools has grown by 30x since October 1."
  • "Since the removal of the original ARTEX repository on October 8, traffic has grown more than tenfold."
  • "As of November 12, ARTEX and its forks make up over 60% of currently known AI penetration testing orchestrators."
  • "In just the last three days, we have seen over 50 framework families attempt to scan Cloudflare customers globally."
  • There are "over 12,000 repositories, with several hundred unique forks" descending from the original. A chart caption counts 204 public forks with their own code changes.

Two caveats come from Cloudflare's own text. The traffic chart behind the growth figures covers "a selected handful of AI-driven pentesting orchestrators", not every tool. The post also does not say what the 60% is a share of: tools, repositories, IP addresses or requests. The values plotted in the charts are not given in the text.

Cloudflare names four other frameworks alongside ARTEX in its top-five table: LuaN1aoAgent, Decepticon, Cairn and PentestGPT-Agent. It describes Cairn as a dispatcher that hands tasks to existing coding agents such as Codex and Claude Code. Cloudflare does not say whether this is the same Cairn that Unit 42 tied to a suspected China-nexus espionage campaign.

Cloudflare's incident work. Cloudforce One says it has been pulled into incident-response cases involving ARTEX and its derivatives "over the last eight weeks". It describes one case, without a date. "Available evidence indicated that the compromise most likely originated through a third-party hyperscaler", meaning a large cloud provider, and the attacker then moved into the customer's Cloudflare environment. "Attackers repeatedly compromised static, non-expiring tokens running inside CI/CD pipelines." Cloudflare counted activity above 69 requests per minute and 400 database-dump requests in an hour. It judges the activity was "lightly guided by the human operator".

The FBI's view. The FLASH describes ARTEX as an open-source, Go-based, multi-agent platform. It can use Anthropic Claude, OpenAI GPT, DeepSeek, GLM, Grok and Kimi models, and attacker-run instances mostly use DeepSeek Flash-family models. It lists target sectors from finance and government to semiconductors and legal services. It says "some intrusions" progressed "from initial targeting to privileged access in under three hours". That claim cites Microsoft's October 21 report and gives no sample size or median. The initial-access methods it lists are almost all old ones: missing authentication on APIs, hard-coded admin passwords in web front-ends, injection flaws, weak admin credentials, stolen VPN configurations, and trust links with partner and vendor tenants.

The European Commission. The Record reports that spokesperson Thomas Regnier said the first signs were detected on October 22. He called it an "AI-generated cyber campaign" and did not name the tool or a motive. A Commission statement said it had "taken the necessary steps in response to a detected cyber campaign", took services offline as "a protective measure", notified affected individuals and was in contact with CERT-EU and the European Data Protection Supervisor. On the gap between detection and disclosure, Regnier said: "There was no hiding whatsoever." He said investigators had found no evidence of a breach affecting staff broadly and knew of no ransom demand. Trade press names the affected systems as Sysper (HR, payroll, leave), MyPMO (medical reimbursement) and translation services.

Technical analysis

Why the takedown backfired. Our assessment: an open-source tool that is popular and simple to copy cannot really be taken down by its author. Removing the repository took away the one maintained version, while the code, the documentation and the publicity remained. The likely result is the pattern Cloudflare describes: hundreds of forks, renamed variants and a paid "PRO" edition, with no single project left for defenders to track. Defenders lost the signal from a single source. Attackers lost almost nothing.

What an orchestrator changes. A classic vulnerability scanner fires a fixed list of checks and stops. An orchestrator puts an AI model in charge of the loop: look, decide, act, look again. Cloudflare's phrase is "a harness performs every step of the OODA loop" (observe, orient, decide, act). In practice this means:

  • Parallelism. Many agents probe different services at once and share findings in a common map of the target.
  • Chaining. A finding in one place, such as a token in a build log, is immediately tried somewhere else, such as a cloud console or a partner's tenant.
  • Persistence. Agents do not get bored. They work through path variations, parameter combinations and old services that a human tester would skip.
  • Scale per operator. Cloudflare notes that "a single operator can target multiple organizations at the same time".

Why controls fail. None of the cases in these sources needed a zero-day. Cloudflare's incident turned on static, non-expiring CI/CD tokens and trust between cloud environments. The FBI's access list is a catalogue of long-known weaknesses. Our own coverage of Barracuda's 44-hour breach and the WorkOS sandbox escape shows the same thing: a forgotten key or internal tool, found quickly. The AI angle is economic, not a new technique. The cost of trying every door on every exposed service has fallen close to the price of the model tokens, and those tokens are sometimes paid for with stolen keys.

What defenders can see. Cloudflare does not explain how it fingerprints orchestrator traffic. It reports deployment artefacts (for example, cloud worker scripts named after ARTEX), aggressive path variations and machine-like request rhythm. The FBI's indicators include outbound connections to AI provider APIs from places that should never make them. Our assessment: the most reliable defender-side signal is not "is this AI?" but behaviour no legitimate user shows. Examples are hundreds of sequential record lookups in one session, a build token used from a new network, or a service account suddenly calling a model API.

What remains unclear

  • What the 60% measures. Cloudflare gives no denominator or method for "currently known" orchestrators. The 30x and tenfold growth figures come from "a selected handful" of tools. The four charts are images whose values are not stated in the text.
  • Counting the ecosystem. Cloudflare says "over 50 framework families" scanned customers in three days, and separately that it is "tracking 50+ AI LLM agents". It does not say whether these are the same set. The FBI's figure of more than 1,600 exposed consoles counts something else again: visible control panels, not traffic.
  • The Commission's ARTEX link. The only basis for the ARTEX link is Politico's anonymous officials. Reports differ on whether ARTEX was used to get in or only to search for weaknesses. The Commission has named neither the tool nor the entry point, has not said what data was read or copied, and has not said why disclosure came four weeks after detection. Reported spokesperson wording also differs slightly between outlets. We could not read Politico's original newsletter.
  • Who was behind it. No source attributes the Commission attack to any actor, and we do not either.
  • Cloudflare's own incident. It is undated and anonymised. Cloudflare does not say whether it is typical of the "eight weeks" of cases it mentions.
  • Vendor framing. Cloudflare's recommendations map closely onto its own products. The underlying advice (rate limits, schema validation, identity-gated access, short-lived credentials) does not depend on any one vendor.

Lessons and what to do

For security and IT operations teams

  1. Assume continuous, automated probing. Plan capacity, alert thresholds and on-call for tooling that never sleeps, not for occasional scans. Rate limits on login, search and export endpoints are now basic protection, not a nice extra.
  2. Kill static credentials in pipelines. Cloudflare's case and Barracuda's both turned on keys that never expired. Move CI/CD to short-lived, workload-identity credentials. Inventory every token older than 90 days and every token belonging to a former vendor.
  3. Shrink the forgotten edge. The FBI's list (partner portals, mobile back-ends, legacy services, management interfaces) is exactly what agents find first. Put admin and publishing interfaces behind identity-aware access, not just IP allow-lists.
  4. Alert on behaviour, not signatures. Watch for sequential object enumeration, bulk exports, and build or service identities used from new networks. Also watch for outbound calls to AI provider APIs from servers that have no business making them.
  5. Vet indicators before blocking. The FBI warns that its IP and domain indicators are historical. Treat a match as a reason to investigate, not as proof.

For AI builders and platform teams

  • If your product exposes an agent console, model gateway or orchestration panel, ship it with authentication on and admin interfaces bound to localhost by default. Exposed consoles are what the FBI is counting.
  • Set per-key spending limits and anomaly alerts with model providers. Stolen keys pay for the attackers' agents.

For leadership

  • A takedown, a ban or a GitHub removal is not a control. Budget for the assumption that capable attack tooling is permanently available.
  • The Commission case shows that a well-resourced institution can lose core staff services for more than a week. Check that payroll, HR and similar "boring" systems are in scope for incident-response plans and backups.
  • Ask your teams one measurable question: how quickly can we find and revoke every credential a departed vendor or a public leak could have exposed? The sources here suggest attackers can act within hours.

Sources

Coverage:

Related on AI Cyber Brief: Amazon and the FBI on ARTEX · Barracuda's 44-hour breach · Unit 42 on Cairn · Microsoft on Storm-3358

Get the daily brief

AI + security signal by email: headlines, a two-line summary, a link. No noise, no spam.

How often