Cloudflare Sees Surge in AI Pentesting Scans, Recommends Modern Security Model

Summary
Cloudflare reports 50+ AI pentesting frameworks scanning customers and a 30-fold rise in requests since Oct. 1. It says exposed CI/CD tokens enabled cross-environment access and recommends WAF, bot defense, identity gating, microsegmentation and schema validation.
Key points
- Cloudflare says more than 50 AI pentesting framework families attempted to scan its customers in three days, with the number of observed IP addresses in the thousands.
- Daily requests from these tools have risen 30-fold since Oct. 1; Cloudflare says ARTEX and its forks accounted for over 60% of known orchestrators as of Nov. 12.
- Cloudflare's Cloudforce One team says it supported multiple investigations involving ARTEX-derived frameworks and observed attackers pivoting from hyperscaler environments into customer Cloudflare environments.
- Cloudflare attributed a recurring access problem to compromised static, non-expiring CI/CD tokens with broad access across environments.
- In one investigation, Cloudflare observed ARTEX using a compromised account-level API token, followed by new Worker scripts, more than 69 requests per minute, 400 database dump requests in an hour, and attempted persistence.
- Cloudflare recommends a six-part defense model: traffic controls and WAF, anomaly detection, bot defense, identity-gated access and microsegmentation, validated API schemas and explicit security controls, and threat intelligence.
Article Details
- Defense Focus
- Detect and contain automated AI-assisted reconnaissance and exploitation, while limiting the impact of compromised credentials through layered access controls and segmentation.
- Detection Methods
- Monitor for automated scanning patterns, unusually high request rates, abnormal request paths, and mechanical bot behavior.
- Watch for unexpected Worker script creation, including names matching the observed artex-*, artex-pgrecon-*, artex-pgw-*, and artex-do-* patterns.
- Investigate spikes in API or database activity, including unusually frequent database dump requests, session-token validation attempts, and persistence attempts.
- Review use of long-lived CI/CD tokens for access across cloud environments and customer accounts.
- Data Sources
- Web and network request activity
- Worker script creation and activity
- API and database request activity
- Session-token validation and CI/CD token use
- Rule Types
- WAF rate-limit and anomaly-detection rules
- Bot behavior detection and challenge rules
- OpenAPI schema-validation rules
- Mutual TLS certificate-validation policies
- Explicit default-deny rules
- Platforms
- Web Application Firewall (WAF)
- Bot Defense and behavioral challenges
- Identity-gated access and Zero Trust controls
- Micro-segmentation
- Defensive Actions
- Route incoming traffic through distributed traffic controls and apply WAF rate limits and anomaly detection.
- Use bot detection and challenges to identify and block automated reconnaissance.
- Require identity-gated access for high-risk applications and apply micro-segmentation to restrict lateral movement.
- Validate incoming requests against OpenAPI schemas, require valid mTLS certificates, and reject requests that fail validation.
- Use hardware FIDO security keys and monitor or restrict long-lived, broadly scoped CI/CD tokens.
- Apply threat intelligence to update detections and blocklists as new infrastructure and techniques emerge.
MITRE ATT&CK
T1078.004 · Cloud AccountsIn an incident, an attacker used a compromised account-level API token to gain access to a customer environment and pivot between cloud environments.T1550.001 · Application Access TokenARTEX used a compromised account-level API token previously used for CI/CD to access the customer's Cloudflare environment.T1595.002 · Vulnerability ScanningThe article reports that AI pentesting frameworks scan customers globally and systematically test exposed services and their configurations.T1648 · Serverless ExecutionAfter gaining access, ARTEX created numerous Worker scripts in the compromised Cloudflare environment.
People
Vendors
Tools
ARTEXLanguage Model (LLM) penetration testing and orchestration frameworks has exploded. Just a few weeks ago, ARTEX — an open source agentic penetration testing system that automates offensive security workflows — was theChiselBuilt-in SSH and Chisel for tunneling and attempted reconnection after disruption
