Cloudflare warns AI pentest frameworks are scaling scans of Internet-exposed infrastructure

Cloudflare warns that automated AI pentest frameworks such as ARTEX are scaling the scanning and exploitation of Internet-exposed infrastructure. In incidents investigated by Cloudforce One, stolen credentials were also used to pivot between hyperscaler environments and Cloudflare.

Overview and technical details

ARTEX and similar frameworks orchestrate AI agents and pentest tools to automatically map targets, find weaknesses, and carry out attack steps. Agents can operate in parallel, build graphs of assets and targets, and then observe, prioritize, and test externally exposed services.

In the investigated incidents, threat actors used static API tokens in CI/CD pipelines that did not expire, had broad access, and were not restricted by IP or rate limits. Cloudflare observed the tokens being used to pivot between hyperscaler environments and Cloudflare.

Affected systems

The warning focuses on Internet-exposed infrastructure and services, web applications, and credentials with broad access in CI/CD pipelines. Static API tokens that do not expire and are not restricted by IP or rate limits warrant particular attention. The supplied information does not identify affected software version branches.

Mechanism and impact

The frameworks can automatically observe and test public-facing services for weaknesses; when combined with stolen credentials, threat actors may use them to pivot between environments. Cloudflare also observed attempts to maintain persistence and authenticate third-party application session tokens.

Observed indicators include newly created Workers named with the patterns artex-*, artex-pgrecon-*, artex-pgw-*, or artex-do-* followed by alphanumeric suffixes; traffic exceeding 69 requests per minute across multiple parts of the infrastructure; and 400 database-dump requests in one hour. These are indicators observed in incidents and do not, by themselves, prove that a system showing them has been compromised.

Remediation and mitigation

1. Prioritize reviewing Internet-exposed applications and services, API tokens, and unusual activity.

2. Review CI/CD credentials; avoid static, non-expiring tokens and broad access. For tokens at risk of exposure, reduce privileges to the necessary level, restrict them by IP and rate, and revoke or rotate them.

3. Apply the six pillars of the Modern Security Model recommended by Cloudflare:

   • Deploy a WAF with advanced rules such as rate limiting.
   • Detect anomalies on customer-facing web applications and deploy Bot Defense.
   • Require Identity-Gated Access for high-risk internal and external web applications.
   • Enforce validated schemas alongside explicit security controls.
   • Use Threat Intelligence to improve attack visibility.

References

• Cloudflare Blog: Adapting to AI penetration testing orchestrators