AI Penetration Testing Agents Put Identity and Lateral Movement at the Center of Defense
- Cargo
- Machine Learning
- Shipped
- Shipper
- Blake Darché
- Readers
- 714
AI testing frameworks can automate reconnaissance and exploitation at scale. The reported incidents show why access limits and segmentation matter.

AI driven penetration testing frameworks are changing the pace and reach of security incidents. They can coordinate existing tools, map target environments and run multiple agents in parallel. In Cloudflare’s observations, more than 50 framework families attempted to scan its customers globally in three days, while the number of IP addresses associated with these systems reached the thousands and continued to rise. The practical concern is the volume of activity these systems can generate, especially when attackers combine automation with exposed credentials.
One prominent framework, ARTEX, was removed by its creator, Autumn-27, on October 8. The creator wrote in Chinese that malicious use ran counter to the tool’s original intent. Its removal did not stop development: the ecosystem that followed included more than 12,000 repositories and several hundred distinct forks with lineage from the original project, along with hundreds of similar orchestration frameworks. Cloudflare reported that daily requests involving AI penetration testing and exploitation tools had grown 30-fold since October 1.
These figures describe activity observed by one network provider, not a complete census of every tool or attack. They still point to a meaningful shift: a single operator can direct automated systems to work across many targets, while agents repeatedly test exposed services and configurations. The systems don’t eliminate the need for human direction in every case, but they can make reconnaissance and follow-on actions faster and more systematic.
What the orchestrators do
AI penetration testing orchestrators connect language model agents with conventional security tools and execution environments. Frameworks such as ARTEX can use Model Context Protocol (MCP) integrations to connect standalone tools, then organize findings and tasks into graphs. Depending on the system, agents may map assets from inside or outside an environment, investigate leads, check evidence, and attempt actions in parallel. Some also accept operator instructions about the goal.
This structure matters because an agent can use the result of one step to guide the next. A discovered asset can become a new task; evidence of a weakness can shape what the system tries afterward. In effect, the framework breaks a complex objective into smaller steps and keeps pursuing promising paths. The same pattern can be useful in authorized testing, but in an intrusion it can increase the speed and breadth of attempts after access is gained.
Cloudflare said it was tracking more than 50 AI agents. Its comparison of five popular, fully autonomous orchestrators—each with more than 1,000 GitHub stars—shows that they differ in how they coordinate work and what they can do after gaining access:
| Framework | Architecture and planning | Execution environment | Post-access features |
|---|---|---|---|
| ARTEX and forks | A planner directs parallel agents. Goals, leads and findings connect to a separate graph of target assets. | Runs on a host or in a container with network and root access; can install or develop tools as needed. | No built-in post-access capability; requires user direction. |
| LuaN1aoAgent | A supervisor, planner and general executor agents use three graphs for goals and blockers, evidence and hypotheses, and target and local context. | Runs in a container with network and user access; can develop tools and install user packages. | Includes SSH and Chisel for tunneling, plus attempts to reconnect after disruption. |
| Decepticon | An orchestrator assigns work to specialist agents; objectives are represented as a graph with prerequisites. | Runs in a Kali container as root and can develop tools as needed. | Includes dedicated skills for post-access activity, Active Directory and cloud pivoting. |
| Cairn | Dispatches work to existing AI harnesses, including Codex, Claude Code and Pi. New fact or lead nodes can trigger the dispatcher. | Runs in a Kali container with root access or directly on a host; can develop tools as needed. | Uses prompt guidance and Kali tools. |
| PentestGPT-Agent | A supervisor and one general executor use a task list containing dependencies, goals and evidence. | Runs on a host and can develop tools as needed. | No built-in post-access capability; requires user direction. |
The operating environments deserve attention. Some frameworks can run with root privileges and network access, while others can develop or install tools during execution. That flexibility can help a system adapt to what it finds, but it also makes the framework’s permissions, isolation and network reach central parts of the risk assessment. The table describes reported capabilities; it does not establish that every deployment uses them in the same way.
Observed growth and credential chaining
As of November 12, ARTEX and its forks accounted for more than 60% of the AI penetration testing orchestrators Cloudflare knew about. Cloudflare reported that traffic had risen by more than tenfold since the original repository was removed on October 8, and by more than 30-fold since October 1. It attributed a sharp increase in unique instances around October 8 to a possible “hydra effect”: removing the original repository may have encouraged more forks and modifications.
The threat becomes more serious when an agentic framework can use credentials that are already exposed. Across multiple incident response investigations involving ARTEX and related systems, Cloudforce One observed attackers entering a cloud service environment and then pivoting into a customer’s Cloudflare environment. The investigations found a recurring weakness: static, non-expiring tokens used in continuous integration and delivery (CI/CD) pipelines had broad access across environments. Once compromised, credentials with that reach can help an attacker move between systems.
One investigation found ARTEX using a compromised account-level API token that had been used for CI/CD. The token had no IP restrictions or rate limits. Available evidence pointed to a compromise that most likely began through a third-party cloud provider. After gaining access, the activity included the creation of many Worker scripts with names such as artex-*, artex-pgrecon-*, artex-pgw-* and artex-do-*. The names corresponded to different observed behaviors and often ended with an alphanumeric identifier.
Investigators observed more than 69 requests per minute across several parts of the customer’s Cloudflare environment, along with 400 database dump requests in one hour. They also saw attempted persistence and attempts to validate session tokens for other third-party applications. Cloudflare assessed that a human operator lightly guided much of the activity. The case therefore shows both sides of the operational picture: automation can sustain a high volume of actions, while a person may still steer the framework.
A six part model for reducing exposure
Cloudflare’s proposed Modern Security Model combines controls at the network edge, application layer, identity boundary and API. Its central operational idea is containment: reduce what an exposed service can reach, restrict which credentials can do, and make suspicious behavior easier to detect.
Control traffic at the perimeter
Network traffic controls and a web application firewall (WAF) can filter abusive requests before they reach applications. Rate limits help constrain request volume, while advanced rules can block unwanted traffic patterns. Cloudflare also points to distributed network controls for volumetric activity and direct IP probing. These protections are most useful when rules reflect the organization’s actual services and expected traffic.
Detect application anomalies
Application defenses need to identify behavior that falls outside expected use, including unusually rapid requests, suspicious path variations and requests for sensitive diagnostic data such as heap dumps. A WAF can combine rate limits with anomaly detection to flag or stop these patterns. The article also recommends security keys and hardware based FIDO keys as part of account protection.
Distinguish automation from people
Automated agents, headless browsers and reconnaissance tools may interact with a service in ways that differ from ordinary users. Bot Defense and challenges can assess behavioral signals and interrupt suspicious automation. Such controls need to account for legitimate automated clients too, so that security measures don’t block expected integrations or accessibility tools.
Hide sensitive services behind identity checks
Internet facing administrative tools, diagnostic endpoints such as /actuator, and remote gateways can expose useful paths into an environment even when they appear locked down. Identity-gated access can make these applications unavailable to unauthenticated visitors and allow access only after identity and device checks. Microsegmentation then limits which systems a user or compromised credential can reach, reducing movement between production, test and development environments.
Validate every API request
API traffic should be checked against explicit schemas and access rules. OpenAPI schema validation can reject requests that don’t match expected structures; mutual TLS (mTLS) certificates can authenticate connections between trusted services. The proposed approach rejects requests that fail validation or lack valid credentials, then ends each rule set with an explicit deny-all rule. This makes the permitted traffic easier to define and audit.
Use threat intelligence to update defenses
Threat intelligence can provide visibility into observed tools and techniques, helping teams update detection rules and watchlists as activity changes. Its value depends on turning information into practical changes: identifying relevant indicators, reviewing exposed services and adjusting controls. Intelligence can complement local telemetry, but it does not replace monitoring an organization’s own systems.
Design for containment
Cloudflare argues that perimeter defenses built around a traditional network demilitarized zone are insufficient for complex environments where credentials and services span cloud and on-premises systems. Its recommendations place identity at the center of access, make internal services unreachable from the public internet where possible, and use microsegmentation to limit the damage a compromised account can cause.
For teams reviewing their exposure, the incident details make two priorities concrete: remove static, long-lived credentials where possible, and narrow the permissions and network reach of tokens that remain. Rate limits and anomaly detection can help control automated request bursts, while identity-gated access and segmentation can restrict what an attacker can do after an initial compromise. As orchestration tools continue to spread, the ability to contain access may matter as much as the ability to block the first request.