AI pentest orchestrators: what Cloudflare is seeing and how to protect yourself
On this page
On November 13, 2026, Blake Darché and Myles Sloan published a warning on the Cloudflare blog: tools that use AI agents to carry out an attack, step by step, are already scanning the company's customers at scale.
We explain what these orchestrators are, what Cloudflare is measuring, how a real incident unfolded and what to do about it. Cloudflare products appear only as examples: the practices apply to any stack.
#What an AI pentest orchestrator is
A pentest (penetration test) means attacking a system, with permission, to find flaws before criminals do. An LLM (large language model) pentest orchestrator automates that work: instead of a person deciding each step, a program coordinates several AI agents.
The steps follow the Cyber Kill Chain, a Lockheed Martin model that describes, in seven steps, what an intruder has to complete to reach their goal. According to Cloudflare, it works like this:
- A planner keeps a graph of goals, leads and findings, linked to a map of the target's assets.
- Parallel agents enumerate, verify, build exploit prototypes and attack.
- Existing pentesting tools plug in as MCP servers (Model Context Protocol, the open standard for connecting models to tools).
- The cycle of observing, orienting, deciding and acting (the OODA loop) runs without a person approving each step.
In practice, the article says, the attacker doesn't need to be an expert, gains speed from parallel agents and can target several organizations at once.
#What Cloudflare is seeing
According to the article, until a few weeks ago ARTEX, an open source agentic pentesting system, was the only framework of its kind available. On October 8, its creator, who goes by Autumn-27, took the repository down, saying malicious use ran counter to its original intent.
The takedown triggered what Cloudflare calls a hydra effect: the copies multiplied. The numbers:
- since October 8, more than 12,000 repositories, hundreds of forks of the original and hundreds of other orchestration frameworks;
- more than 50 framework families trying to scan Cloudflare customers in just three days;
- thousands of IP addresses running these tools, a number that grows every day;
- daily request volume from these tools up 30x since October 1, and more than 10x since the takedown;
- as of November 12, ARTEX and its forks made up more than 60% of known orchestrators.
#The five most popular
Cloudflare compared five, each with more than 1,000 stars on GitHub:
- ARTEX and forks: a planner and parallel agents.
- LuaN1aoAgent: agents to supervise, plan and execute, with SSH and Chisel tunnels for the post-breach phase.
- Decepticon: an orchestrator that hands tasks to specialist sub-agents, with skills for Active Directory and the cloud.
- Cairn: passes the work to existing AI assistants such as Codex, Claude Code and Pi.
- PentestGPT-Agent: one supervisor and a single executor agent.
#A real incident: the forgotten CI/CD key
Cloudforce One, Cloudflare's incident response team, assisted a customer attacked with ARTEX. Initial access most likely came through a large third-party cloud provider. From there, the attacker jumped into the customer's Cloudflare environment with an account-level API token previously used in CI/CD (the pipeline that tests and ships code). The token had no IP restriction and no rate limit.
The investigators found:
- many new Workers scripts (Workers is Cloudflare's serverless platform) with names such as
artex-*,artex-pgrecon-*andartex-do-*; - more than 69 requests per minute across several parts of the infrastructure;
- 400 database dump (full copy) requests in one hour;
- attempts at persistence, to keep access, and at validating session tokens for other applications.
Much of this, according to Cloudflare, was only lightly guided by a human operator.
#The usual cause, now more dangerous
Over the last eight weeks, Cloudforce One handled several cases involving ARTEX and its derivatives. The common cause was credential hygiene: static, non-expiring tokens used in CI/CD pipelines with broad access to several cloud environments. The article calls this identity-chained exploitation: one stolen credential opens the next door.
A leaked key is nothing new. What changed is the pace: with LLMs, each stage of the attack is validated and feeds the next, and, according to Cloudflare, the odds that a leaked key gets used to move further into the environment rise sharply.
#The six layers Cloudflare proposes
The article calls its answer the Modern Security Model and uses the metaphor of a house. No new product was announced.
- Fortify the front gate: traffic control and a web application firewall (WAF) with rate limiting.
- A security desk in the foyer: anomaly detection and hardware security keys based on the FIDO standard.
- Spot fake visitors: detect bot-like behavior (at Cloudflare, Bot Management and Turnstile).
- Close the back door: internal systems reachable only with a verified identity (Access) and micro-segmentation to contain lateral movement.
- Inspect deliveries: validate requests against the OpenAPI schema (API Shield), require mTLS (the client also presents a certificate) and deny whatever hasn't been allowed.
- Neighborhood watch: threat intelligence to learn early about new techniques.
#Applying this to any stack
- Short lifetimes and minimal scope. One token per purpose and per environment, with only the permissions it needs and an expiration date. The Cloudflare dashboard, for example, lets you set an expiration and restrict a token by IP.
- OIDC in CI instead of static keys. With OpenID Connect, the pipeline asks the cloud provider for a temporary credential. In GitHub Actions it is valid for a single job, and no long-lived secret is stored.
- IP restrictions and rate limits. If only CI uses the token, it should only work from CI's IP addresses. No legitimate automation needs 400 database dumps in an hour.
- Deny by default. Allow known routes and formats and reject the rest.
- Watch the volume. Alerts for request spikes, bulk exports and new resources with odd names.
- MFA with a hardware key for every admin account.
#What to review this week
- List the API tokens and keys in CI/CD, on servers and on your team's machines, and revoke the ones nobody can explain.
- Give the remaining ones an expiration date and the smallest possible scope.
- Replace static CI keys with OIDC wherever your provider supports it.
- Check whether any admin panel is exposed to the internet.
- Turn on alerts for abnormal volume and for resource creation.
#What this changes for teams building products and agents
At MEYJE we build AI agents for clients, and the article has an uncomfortable side for anyone doing that: attack orchestrators use the same architecture as legitimate agents, with a planner, parallel agents and tools over MCP. The difference lies in intent and permissions.
Three consequences are now part of how we design:
- Every agent is an identity. Its own credential, short-lived and with minimal permissions, never the account's master key.
- An agent's tools are an API. Schema validation, rate limits and deny-by-default apply to whatever the agent exposes.
- Abnormal volume is a signal. A legitimate agent has a predictable usage pattern. Alerting when it strays protects against attacks and against the agent's own mistakes.
If you want to put AI agents to work at your company with security built in from the start, talk to us.