Published
- 4 min read
By Allan D - Editor, AI Security Wire
Cloudflare: 50+ AI Pentest Frameworks Scanning, ARTEX 60% of Tracked
On November 13, Cloudflare’s Cloudforce One team published its view of a fast-growing problem: AI-driven penetration-testing tools pointed at the open internet. Its headline claim is that in the last three days it saw over 50 framework families attempt to scan Cloudflare customers globally, from thousands of IP addresses, with the count still rising.
All figures below are Cloudflare’s own network telemetry and incident-response findings from a single vendor post, with no independent confirmation that we found. We could not verify them independently, and Cloudflare sells products in the categories it recommends.
The scale Cloudflare describes
ARTEX, an open-source agentic penetration-testing system, was the only such framework a few weeks ago, per Cloudflare. Its creator, Autumn-27, took it down on October 8, writing that malicious use “runs entirely counter to its original intent”. Cloudflare says the ecosystem has grown since: over 12,000 repositories, several hundred unique forks with lineage from the original, and hundreds of other agentic orchestration frameworks created after October 8.
Its numbers for traffic: daily requests from AI pentesting and exploitation tools up 30x since October 1, and more than tenfold since the takedown. As of November 12, ARTEX and its forks are over 60% of the AI pentest orchestrators Cloudflare tracks. It calls the post-takedown surge a likely “hydra effect”. That is Cloudflare’s interpretation, not something we can test.
What the tools do
Cloudflare tracks 50+ of these agents and lists five popular ones (each over 1,000 GitHub stars, it says): ARTEX and its forks, LuaN1aoAgent, Decepticon, Cairn and PentestGPT-Agent. Its comparison table shows common traits. Most run a planner or supervisor over executor agents, keep goals and findings in a graph, run in a container or on the host with network access, and can write or install tools as needed. Decepticon has dedicated Active Directory and cloud pivoting skills. LuaN1aoAgent has built-in SSH and Chisel tunnelling with reconnection attempts. Cairn dispatches work to existing harnesses (Codex, Claude Code, Pi).
Cloudflare’s argument is that the limit is no longer human skill. Parallel agents work through observe, orient, decide and act loops at the speed of inference, one operator can hit several organisations at once, and small configuration mistakes get found systematically.
The incident detail
The most concrete part is a response case. Cloudflare says it observed ARTEX gaining access to a compromised account-level API token, previously used for CI/CD, with no IP restrictions or rate limiting. The evidence indicated the compromise most likely originated through a third-party hyperscaler. The intruder created many new Worker scripts named artex-, artex-pgrecon-, artex-pgw-* and artex-do-*, made more than 69 requests per minute across several parts of the customer’s Cloudflare setup, and issued 400 database dump requests in an hour. Cloudflare also saw attempted persistence and attempts to validate session tokens for other third-party apps. It believes a human operator only lightly guided much of this.
Across several incidents over eight weeks, it says, the common root cause was static, non-expiring tokens in CI/CD pipelines that let attackers pivot from a hyperscaler into a customer’s Cloudflare environment.
What defenders can take from it
Cloudflare’s own framing, the “Modern Security Model”, has six pillars: WAF with rate limits, anomaly detection, bot defence, identity-gated access for high-risk apps, validated API schemas with mutual TLS, : the parts that stand apart from any vendor are the credential and exposure ones. Replace long-lived CI/CD tokens with short-lived ones, bind API tokens to source IPs, put admin and diagnostic endpoints behind identity, and segment so one stolen credential does not reach production, test and development at once.
For earlier coverage of this tooling, see Unit 42’s defender guidance on ARTEX, Check Point’s research on ARTEX operators and Unit 42 on Cairn.
References
Frequently Asked Questions
- What did Cloudflare report on November 13, 2026?
- Cloudflare's Cloudforce One team said that in the last three days it saw over 50 AI penetration-testing framework families attempt to scan Cloudflare customers globally, with the number of IP addresses running them in the thousands and still rising. It also said ARTEX and its forks make up over 60% of the AI pentest orchestrators it currently tracks, as of November 12.
- Did the ARTEX takedown stop the tool?
- Not according to Cloudflare. It says the original repository was removed by its creator on October 8 and that traffic has since grown more than tenfold, with a likely 'hydra effect' as forks multiplied. It counts over 12,000 repositories and several hundred unique forks with lineage from the original. These are Cloudflare's own figures.
- What credential failure did Cloudflare highlight?
- Cloudflare says that in multiple incidents attackers compromised static, non-expiring tokens running in CI/CD pipelines, which gave broad access across an organisation's hyperscaler infrastructure. In one case it saw an account-level API token with no IP restrictions or rate limiting used to create new Worker scripts named artex-*.
- What does Cloudflare recommend?
- It proposes a six-part 'Modern Security Model': a WAF with rate limits, anomaly detection, bot defence, identity-gated access with micro-segmentation, validated API schemas with mTLS, and threat intelligence. Several of these map to Cloudflare products, so weigh the framing accordingly. Our assessment: the token and segmentation advice applies on any vendor's stack.