On November 13, 2026, Cloudflare published data from its own network on AI penetration testing orchestrators: systems that split the work of a penetration test across several agents running in parallel, which map assets, test hypotheses and try to exploit what they find. According to the report, daily requests from these tools grew 30x since October 1. In the incident described in the post, the way in was an API token used in CI/CD, with no expiration date and no IP restriction.
What Cloudflare measured
The starting point is ARTEX, an open source agentic pentest system. Until a few weeks before the report, it was the only one of its kind. On October 8, its author took the repository down and wrote that malicious use of the tool ran counter to its original intent. The takedown did not stop it from spreading. The report's figures, with data through November 12:
- more than 12,000 related repositories, with hundreds of forks derived from the original (204 with their own code changes);
- traffic from these orchestrators grew more than tenfold since the October 8 takedown;
- ARTEX and its forks account for more than 60% of known orchestrators;
- more than 50 tool families tried to scan Cloudflare customers in just three days, from thousands of IP addresses.
The report names admin panels, remote access gateways and exposed diagnostic endpoints as targets, with /actuator as an example, plus path tricks to get around access rules and attempts to download heap dumps (copies of the application's memory, which can contain secrets).
The incident started with a pipeline token
In investigations over the past eight weeks, Cloudflare's incident response team (Cloudforce One) describes attackers who first got into a cloud environment at another, unnamed provider and moved from there into the customer's Cloudflare account. The cause, in the report's words, was a common credential hygiene failure: static, non-expiring tokens running in CI/CD pipelines.
In one case, ARTEX used an account-level API token that served CI/CD, with no IP restriction or rate limit. With it, it created Workers with names such as artex-pgrecon-* and artex-do-*, went above 69 requests per minute across several parts of the infrastructure and made 400 database dump requests in one hour. It also tried to establish persistence and to validate session tokens for third-party applications.
Cloudflare notes that much of this activity was lightly guided by a human operator. The tool shortens the work, but the way in was a credential that was already there, with more reach than its task required. The permissions on that token set how much damage was possible, as in the case discussed in The Vercel incident and the cost of an integration with too much access.
Who is most exposed
Any team that deploys to Cloudflare from a pipeline (GitHub Actions, GitLab CI or similar) and created the token once, with broad permissions, and never touched it again. The risk grows when the token covers the whole account instead of one zone or one Worker, and when the same secret is copied across several repositories or environments.
The pattern applies outside Cloudflare too. Cloud keys, package registry tokens and database credentials stored as CI secrets share the same profile: wide reach, little monitoring and no expiration date.
Sites with an admin panel or diagnostic endpoint open to the internet also show up in the report as direct scanning targets. With thousands of IPs running these tools, the volume no longer depends on someone choosing your site.
What to review this week
There is no formal deadline, because Cloudflare did not change any product or default. Given the volume described, I would do the token review this week.
- List your Cloudflare API tokens (user tokens under My Profile and account tokens under Manage Account) and mark which ones live in pipelines. The documentation allows editing a token after it is created.
- Cut each token down to what the pipeline does. A Worker deploy rarely needs permissions over the whole account or DNS edit rights on every zone.
- Fill in the two optional fields that were empty in the incident: "Token expiration" and "Client IP Address Filtering". IP filtering works best with a self-hosted runner or fixed network egress; on hosted runners the egress IP varies, and expiration with scheduled rotation is the more practical control.
- Take diagnostic endpoints and admin panels off the public internet, or put them behind identity-based login.
- Check the account audit logs for Workers or resources created outside the normal pipeline. The
artex-prefixed names Cloudflare cites are an example, not a reliable signature.
The report also recommends a WAF with rate limiting, bot defense and schema validation for APIs. Those are reasonable controls, but all of them are Cloudflare products; the item that requires no purchase is the pipeline credential.
What the report does not say
The numbers come only from traffic that crosses Cloudflare's network, and the post does not explain how it attributes a request to a specific orchestrator. One IP can run several tools, so the address count is not a count of attackers.
The report does not name the cloud provider the attackers came through, or how many customers were compromised. It also does not say how autonomous the agents really were in the investigated cases, beyond noting human guidance. The view that the ARTEX takedown accelerated its spread is Cloudflare's interpretation, not a measurement.
If your company is weighing something similar, reach me on LinkedIn or WhatsApp.