AI pentest orchestrators: Cloudflare sees 30x more traffic

AI pentest orchestrators are now background traffic on the internet. Cloudflare's threat team, Cloudforce One, reported on November 13 that daily requests from AI penetration testing and exploitation tools grew 30x since October 1. In the incidents it investigated, the attackers did not use a clever exploit. They used a static API token from a CI/CD pipeline that had no expiry.

What actually happened

Cloudflare published Adapting to AI penetration testing orchestrators. The numbers from the post:

• 30x. Growth in daily requests from the tools Cloudflare tracks, since October 1.
• More than 50 framework families tried to scan Cloudflare customers in a three-day window.
• More than tenfold. Traffic growth since October 8, when the creator of ARTEX removed the original repository. Cloudflare says the takedown likely caused a "hydra effect."
• Over 60%. The share of known orchestrators that are ARTEX or its forks, as of November 12. Cloudflare counts over 12,000 repositories and several hundred unique forks with lineage from the original.
• Thousands of IPs run these frameworks. The post gives no exact count.

These tools connect AI agents to standard pentest tools. They map a target onto a graph, then parallel agents enumerate, verify and attack. Some include tunneling and cloud pivoting. We covered exposed ARTEX instances and the Storm-0940 campaign earlier.

The incident detail is the useful part. Cloudforce One says the attackers compromised static tokens with no expiry, again and again, in CI/CD pipelines. In one case, ARTEX used an account-level Cloudflare API token with no IP restriction and no rate limit. It created Worker scripts with artex- names, sent 400 database dump requests in one hour, and tried to validate session tokens for other applications. Cloudflare says the token most likely leaked through a third-party cloud provider, and that a person lightly guided the activity.

The post does not say how many customers were affected. It is also a product pitch: its six-pillar "Modern Security Model" maps to Cloudflare products.

Why AI pentest orchestrators matter for your business

The skill barrier is gone. One operator with a free framework can now probe many companies in parallel, and the framework finds the small mistakes: the forgotten token, the admin panel, the debug endpoint. A small business is not too small to scan when the scan costs nothing.

You do not need a new product to respond. Do this:

1. List your long-lived tokens. Cloudflare, GitHub, AWS, Vercel, Stripe. Look in CI/CD secrets first.
2. Give each token an expiry and a narrow scope. One zone, one job. No account-level tokens in a pipeline.
3. Restrict by IP where the provider lets you. The token in Cloudflare's incident had no restriction.
4. Rate limit and default-deny. Block /actuator, heap dumps and admin paths at the edge.
5. Put internal tools behind identity. An admin panel on the open internet will be found.
6. Watch for new resources. An unknown Worker script or cloud function is an alarm.

Sources: Cloudflare: Adapting to AI penetration testing orchestrators: The Modern Security Model.