AI penetration-testing agents raise the pressure on enterprise defenses
Autonomous penetration-testing frameworks are increasing the volume and pace of probing against internet-facing organizations, according to observations reported by Cloudflare. The company said it saw more than 50 framework families attempt to scan its customers globally during a three-day period, with thousands of IP addresses associated with the tools and that number continuing to rise.
The activity reflects a change in how offensive security workflows can be assembled and run. Open-source systems such as ARTEX use large language model (LLM)-driven agents to coordinate reconnaissance and other testing tasks, sometimes connecting established security tools through Model Context Protocol (MCP). For defenders, the concern is not simply automated scanning: frameworks can operate multiple agents in parallel, validate findings and use results to guide subsequent actions.
From one project to a growing ecosystem
Cloudflare said ARTEX’s creator removed the original project on October 8, stating that malicious use ran counter to its intent. The takedown did not end development. Cloudflare reported more than 12,000 repositories associated with the wider ecosystem, including several hundred distinct forks with lineage from ARTEX, alongside hundreds of other agentic orchestration frameworks created since October 8.
As of November 12, ARTEX and its forks accounted for more than 60% of the AI penetration-testing orchestrators Cloudflare knew about. The company said requests attributed to these tools had risen 30-fold since October 1 and more than tenfold since the original repository was removed. It also observed over 50 AI agents and identified five prominent orchestrators with more than 1,000 GitHub stars. These measurements describe Cloudflare’s own network observations, rather than a complete census of activity across the internet.
The frameworks differ in their architectures and capabilities. Some use planner and executor agents or graphs to track targets, findings and tasks; others dispatch specialist agents or connect to existing AI coding harnesses. Depending on the tool, execution may take place in a container or directly on a host, with varying levels of access and built-in post-compromise functions. That variation makes it difficult for defenders to rely on one signature or a single tool-specific response.
Credential exposure can turn probing into broader access
Cloudflare’s Cloudforce One team said it had supported multiple incident investigations involving ARTEX and related frameworks over an eight-week period. In incidents described by the company, attackers used exposed credentials to move from an initial foothold in a hyperscaler environment into a customer’s Cloudflare environment. Investigators identified static, non-expiring tokens used in CI/CD pipelines as a recurring credential-hygiene weakness. The tokens provided broad access across environments.
In one investigation, Cloudflare observed ARTEX using a compromised account-level API token previously used for CI/CD. The token had no IP restrictions or rate limiting, and available evidence indicated that the compromise most likely originated through a third-party hyperscaler. The company reported the creation of numerous Worker scripts with names including “artex-*” and related variants, more than 69 requests per minute across different parts of the customer’s infrastructure, and 400 database dump requests in one hour. It also observed attempted persistence and efforts to validate session tokens for other third-party applications.
Those observations underline an operational risk: a credential that reaches multiple systems can give an automated workflow opportunities for lateral movement. Cloudflare said some of the more complex activity involved chaining stolen credentials. Organizations should therefore treat exposed API keys and long-lived tokens as potential paths into connected environments, not as isolated application secrets.
Controls should combine access, inspection and containment
Cloudflare recommends a layered approach it calls the Modern Security Model. Its six pillars focus on reducing exposed attack surface and limiting the consequences of successful access:
-
Use traffic controls and a web application firewall (WAF), including rate limits and rules for suspicious request patterns.
-
Monitor customer-facing applications for anomalies, such as unusual request rates, paths or diagnostic-data access.
-
Apply bot defenses to identify automated behavior that may not be distinguishable through basic request signatures alone.
-
Place high-risk internal and external applications behind identity-gated access, and use microsegmentation to restrict movement between systems.
-
Validate API requests against defined schemas, authenticate trusted connections with controls such as mutual TLS where appropriate, and reject traffic that does not meet explicit requirements.
-
Use threat intelligence to improve visibility into emerging tools and activity.
The recommendations emphasize both prevention and containment. Rate limiting and anomaly detection can make large-scale probing harder, while identity checks can remove administrative interfaces from public reach. Microsegmentation can narrow what a compromised account or token can access. Schema validation and default-deny rules can also help ensure that applications accept only expected requests.
For enterprise teams, the immediate priorities are to inventory long-lived credentials, reduce their permissions and scope, and remove or rotate tokens that are no longer needed. CI/CD secrets should be protected and monitored, with access constrained to required systems and contexts. Organizations should also review exposed administrative and diagnostic endpoints, application-level rate controls, and the boundaries between development, test and production environments. As agentic frameworks proliferate, resilience depends not only on detecting a particular scanner but on making each step of an attack harder to complete and limiting the impact when credentials are compromised.