Summary: Open-source AI penetration testing orchestrators went from a single framework — ARTEX — to dozens of families in roughly a month, and Cloudflare’s threat research now recommends a six-pillar defense it calls the Modern Security Model. Cloudflare measured daily requests from these tools growing 30x since October 1 on its own network. Yet every breach it describes in detail entered through a static, unrestricted CI/CD token — a problem an afternoon of credential hygiene can shrink.
What an LLM pentest orchestrator actually does
These frameworks wire standalone pentesting tools into an LLM planner over the Model Context Protocol (MCP). ARTEX and its forks map a target’s assets and environments onto a graph, then run multiple independent parallel agents that enumerate, verify, prototype, and attack at the speed of inference. Each step of the Cyber Kill Chain gets validated and fed into the next stage, which Cloudflare says turns compromise into a hill-climbable problem — the kind of search agents excel at. One operator can point the harness at multiple organizations at once, which multiplies the blast radius.
Cloudflare is tracking 50+ of these agents. Its top five — ARTEX (and forks), LuaN1aoAgent, Decepticon, Cairn, and PentestGPT-Agent — each have 1k+ GitHub stars and run fully autonomously, completing the full observe-orient-decide-act loop without a human at each step. They differ most after initial access: Decepticon ships dedicated Active Directory and cloud pivoting skills, LuaN1aoAgent tunnels over built-in SSH and Chisel and tries to reconnect after disruption, while ARTEX builds in nothing and needs operator instruction.
The takedown that spawned hundreds of forks
ARTEX’s creator, Autumn-27, removed the original repository on October 8, writing in Chinese that malicious exploitation of the tool ran counter to its original intent. The takedown backfired. Cloudflare counted over 12,000 repositories in the ecosystem, several hundred unique forks carrying lineage from the original, and hundreds of additional similar frameworks created since October 8 — a “hydra effect,” in its words. As of November 12, ARTEX and its forks make up over 60% of currently known AI penetration testing orchestrators. Traffic on Cloudflare’s network grew more than tenfold since the takedown and more than 30x since October 1. In just the three days before publication, Cloudflare saw over 50 framework families attempt to scan its customers globally, from IP addresses numbering in the thousands.
The breaches ran on boring credentials
Over the eight weeks before November 13, Cloudflare’s Cloudforce One team joined multiple incident response investigations involving ARTEX and ARTEX-derived frameworks. The common root cause was not exotic AI tradecraft: attackers repeatedly compromised static, non-expiring tokens inside CI/CD pipelines — tokens with no IP restrictions and no rate limiting — that granted broad, cross-environment access across hyperscaler infrastructure, then pivoted directly into the victims’ Cloudflare environments.
In one investigation, a compromised account-level API token let ARTEX create Worker scripts named artex-, artex-pgrecon-, artex-pgw-, and artex-do-, generate more than 69 requests per minute across the victim’s infrastructure, and fire 400 database dump requests in an hour, alongside persistence attempts and validation of session tokens for third-party applications. Cloudflare notes a human operator lightly guided much of this activity.
The Modern Security Model, in one pass
Cloudflare’s prescription has six pillars: fortify the front gate with a Cloudflare-style WAF and advanced rules such as rate limits; run anomaly detection on every customer-facing web application; deploy Bot Defense to flag automated agents that don’t behave like people; put Identity-Gated Access on all high-risk external and internal web applications; enforce validated schemas with explicit security controls; and use threat intelligence for visibility into attacks. The source post develops each pillar through an estate-security metaphor and goes further than we cover here.
Our take: Act now — but spend the first afternoon on credentials, not on the six-pillar package. Every incident Cloudflare documented hinged on a static, non-expiring CI/CD token with no IP restrictions and no rate limiting; expiry dates, scopes, and IP allowlists are cheap, reversible fixes a small team can ship without new vendors or platform engineers. Teams whose applications are internal-only behind identity-gated access can reasonably skip the urgency for now. Two caveats temper this. Every traffic number comes from Cloudflare’s own network and its own incident responses — a vendor describing a threat its products defend against — so treat the 30x figure as a prompt to check your own logs, not as an internet-wide measurement. And nothing shown here demonstrates these frameworks defeating a target that already had scoped tokens and rate limits in place; the documented chains exploited the gap before the pillars. If Cloudflare or an independent responder documents a breach of a target with expiring, scoped, rate-limited credentials, that would change our answer.
