A summary and commentary. Source: Cloudflare blog, Adapting to AI penetration testing orchestrators: The Modern Security Model, published 13 November 2026. The figures below are Cloudflare’s own and come from its network telemetry; I have not verified them independently.
What changed
A few weeks before the report, ARTEX, an open-source agentic penetration-testing system, was the only one of its kind. In the three days before the report, Cloudflare saw more than 50 framework families scanning its customers worldwide. The original repository was taken down on 8 October. The ecosystem did not disappear: it split into more than 12,000 repositories and several hundred forks with ARTEX lineage, and hundreds of similar orchestrators appeared after that date.
According to the report, daily requests for AI pentest and exploitation tools grew about 30 times since 1 October. ARTEX and its forks account for over 60% of the known orchestrators.
What an orchestrator does
These frameworks use AI agents to run the stages of the kill chain at scale. They plug in existing MCP servers and standalone pentest tools, and they plan work as a graph of goals, leads and findings. Several agents enumerate, verify, prototype and attack in parallel.
Cloudflare’s main point is that this turns compromise into a search problem that agents handle well. A human no longer limits the speed. One operator can hit many targets at once, and the framework tests exposed services and their configuration systematically, so small mistakes get found.
The report compares five projects by architecture: where the planner runs, whether the agent has a built-in execution environment, and whether it includes post-exploitation tooling. The differences matter for defenders. A framework with a Kali container and pivoting skills is a different threat from one that runs on a host and waits for instructions.
What the attacks look like
The most useful part of the report is the identity-chained case. Attackers got into a hyperscaler environment and moved into a customer’s Cloudflare account. The root cause was ordinary: long-lived, non-expiring tokens in CI/CD pipelines, with broad access across environments and no IP restrictions or rate limits.
One investigated case shows the pattern. Many new Worker scripts appeared with names matching their behaviour, such as artex-*. The activity included more than 69 requests per minute across several parts of the infrastructure and 400 database dump requests in an hour. Cloudflare believes a human lightly guided much of it, but the speed and breadth after the initial compromise are the real lesson.
The recommended model
Cloudflare’s answer is what it calls the Modern Security Model. The idea is to stop relying on a single perimeter and to layer controls, each of which answers one part of the problem:
- Gate the edge. Route all traffic through distributed filtering, so that only legitimate requests reach the origin. This covers volumetric floods and direct IP probing.
- Detect anomalies. Automated tools send malformed paths and unusual request patterns. Block such behaviour immediately, not after a human reviews it.
- Separate bots from people. Scripts and headless browsers have a mechanical cadence. Bot detection and challenges catch it.
- Remove the back door. Put admin panels, diagnostic endpoints and remote gateways behind identity and device checks, and micro-segment the network so a stolen credential cannot reach production, test and dev together.
- Validate every request. Check incoming requests against an OpenAPI schema, require client certificates for sensitive paths, and end every rule set with an explicit deny-all.
- Use threat intelligence. Update watchlists with new C2 infrastructure and tunnelling tools before an attack arrives.
The report also says that the classic DMZ is not enough for modern architectures, and that legacy and on-prem systems should not be reachable from the internet at all.
My take
The most practical lesson is not about AI. It is about tokens. The compromise that made this scale was a long-lived credential in a CI/CD pipeline, and an agent is good at finding and reusing exactly that kind of secret. If you run automation, make its credentials short-lived, scoped and IP-restricted, and treat a leaked token as an incident by default.
The second lesson is that the cost of finding weaknesses has dropped. A defender can now run the same kind of tools against their own perimeter before someone else does. Regular automated self-testing is no longer an expensive annual exercise, and it belongs in the normal development loop.
See the original report for the full architecture comparison, the traffic charts and the complete list of recommendations.