Autonomous penetration testing is the use of software, usually driven by AI agents, to carry out the steps of a security test with little or no human direction. A person states a goal, such as “find a way into this environment,” and the system decides what to scan, which weakness to try, how to raise its access and what data to collect, adjusting as it learns.
A penetration test is an authorized, simulated attack against a system to find weaknesses before real attackers do. Traditionally it is done by skilled people over days or weeks. Autonomous tools aim to do the repetitive and exploratory parts at machine speed, and run continuously instead of once a year.
The same capability that helps defenders helps attackers. In November 2026, the FBI warned that an open-source autonomous penetration testing framework called ARTEX was being used in real attacks. This page explains how these tools work, how they differ from automated scanning and from human testing, what the ARTEX advisory tells us, the legitimate uses, and how to defend against an adversary that moves at machine speed.
How it differs from related terms
| Approach | What it does | Human involvement |
|---|---|---|
| Vulnerability scanning | Checks systems against a list of known flaws and reports matches | Low. Runs on a schedule and produces a list. |
| Manual penetration testing | Skilled testers chain weaknesses to reach a goal | High. Creative, slow, and limited in scale. |
| Breach and attack simulation | Replays known attack techniques to test detection controls | Medium. Scripted scenarios. |
| Automated penetration testing | Runs scripted exploit chains against a target | Low to medium. Follows predefined steps. |
| Autonomous penetration testing | AI agents decide the next step based on what they find | Low. Goal-driven and adaptive. |
| Red teaming | A team simulates a real adversary against people, process and technology | High. Broader than technical weaknesses. |
The key difference is adaptation. A scanner reports a flaw. An autonomous agent tries it, sees the result, picks a next step, and tries something else if it fails.
How it works
Most autonomous frameworks follow a similar structure.
- Goal. The operator specifies a target and an objective, such as obtaining administrator access or finding sensitive data.
- Planning. A large language model (LLM), which is the type of AI that powers chat assistants, breaks the goal into tasks.
- Specialized agents. Separate agents handle reconnaissance, vulnerability discovery, exploitation, privilege escalation and data collection. Each can call ordinary security tools and write its own scripts.
- Feedback loop. Output from each step goes back to the model, which decides what to do next. When one path fails, it tries another.
- Reporting. The framework summarizes what it found and what it did.
Frameworks usually do not contain their own model. They connect to external LLMs. The FBI reported that ARTEX can use models from Anthropic, OpenAI, DeepSeek, GLM, Grok and Kimi.
The ARTEX advisory
On November 13, 2026, the FBI released a FLASH advisory on ARTEX, as reported by BleepingComputer. According to that report:
- ARTEX is an autonomous penetration testing framework written in Go, first published on GitHub in July 2026 and removed by its developer on October 8 after reports of abuse. It remains available through forks, archived copies, localized versions and a commercial edition.
- The FBI said multiple threat actors have used it since at least August 2026 against organizations in finance, government, technology, defense, healthcare and other sectors.
- Some intrusions progressed from initial targeting to privileged access in under three hours.
- The framework lets actors with limited technical skill run complex, multi-target operations at the same time, and it is especially effective at finding weaknesses in overlooked systems such as partner portals, employee applications, legacy services and third-party integrations.
- Researchers identified more than 1,600 publicly exposed ARTEX consoles as of early November.
We cover the advisory in detail in The FBI warns that ARTEX can take an attacker from first contact to privileged access in under three hours, and a related incident in Barracuda’s cloud breach.
Legitimate uses
Used with authorization, autonomous testing has real value.
- Continuous coverage. Annual tests miss what changes in between. Autonomous tools can test after every significant change.
- Breadth. Agents can check every internet-facing asset, not only the ones a tester had time to examine.
- Validation. They can confirm whether a flagged weakness is actually exploitable, which helps prioritization. See exposure management.
- Finding the forgotten. They are good at turning up the overlooked systems that scanners and inventories miss.
- Reducing cost for smaller teams who cannot afford frequent manual tests.
Responsible use needs written authorization, a defined scope, safe-guards against damaging actions, careful handling of any credentials or data the tool collects, and human review of the results. The testing standards in NIST SP 800-115 and the OWASP Web Security Testing Guide remain good references for planning and scoping.
The risks
- Misuse. The same framework in an attacker’s hands is an intrusion tool, as the FBI advisory shows.
- Unintended damage. An agent can take actions the operator did not expect. Public reporting this year describes AI agents escaping test sandboxes and reaching real systems during safety testing. Treat any autonomous tool as capable of leaving its box.
- Over-trust in results. A clean report does not prove a system is safe. Agents can miss business-logic flaws and can fail silently.
- Data handling. The tool may collect credentials and sensitive data and send prompts containing them to an external model provider.
- False confidence in coverage. The tool tests what it can reach. It does not know about assets outside its scope.
How to defend against autonomous attackers
Autonomous attackers rely on the same weaknesses that human attackers do, found faster. The defense is to have fewer of those weaknesses at rest, and to react quickly to the rest.
Remove the easy finds
- Know your internet-facing assets. Overlooked portals, test systems and old services are what the tools find first. See attack surface management.
- Close default and weak authentication. Autonomous agents try default and leaked credentials quickly.
- Patch by exposure and exploitation. See KEV and EPSS and zero-day response.
- Keep secrets out of code. Agents scan public repositories for keys. See secrets sprawl.
Limit what a foothold can do
- Apply least privilege and remove standing privileges. See zero standing privilege.
- Use short-lived credentials for machines and pipelines, and retire access when work ends. See orphaned credentials.
- Make lateral movement hard. See lateral movement in the cloud.
See and respond at speed
- Alert on machine-like behavior: bursts of requests, many access-denied errors, rapid enumeration, and privilege changes. The FBI recommended watching for bursts of automated requests and unexpected credential or privilege changes.
- Apply rate limits to sign-in, token and API endpoints.
- Automate the safe responses, such as disabling a credential or ending a session, so the response does not wait for a person.
- Use published indicators. The FBI advisory includes indicators of compromise, and a researcher released an open-source detection pack for ARTEX activity. Treat network indicators as short-lived.
How Cloudanix helps
Cloudanix is not a penetration testing product. It does not run attacks, and it does not replace authorized testing or red teaming. It reduces what an autonomous attacker can find and use, and helps you see the paths it would take.
- Fewer standing privileges to chain. Cloud JIT replaces standing access to AWS, Azure and GCP with time-boxed grants that revoke themselves. Database JIT, VM JIT and Kubernetes JIT do the same for databases, servers and clusters, and Agentic JIT and Coding Agent JIT give pipelines and AI agents credentials that last for a task.
- Findings ranked by what an attacker could actually reach. CSPM recomputes severity per asset using exposure, environment, data sensitivity and identity, and the attack path view shows how a foothold connects to sensitive data. CIEM and non-human identity visibility show which identities and keys would be worth stealing.
- Catching secrets before they are public. Code Security scans pull requests and full git history for hardcoded credentials.
- Spotting machine-speed behavior. Cloud UEBA and cloud detection and response look for identities acting unlike themselves.
An honest note: Cloudanix performs static analysis, dependency analysis and cloud posture assessment with runtime context. It is not a dynamic testing tool and it does not validate exploitability by attacking your systems.
Further reading
- NIST SP 800-115: Technical Guide to Information Security Testing and Assessment
- OWASP Web Security Testing Guide
- BleepingComputer: FBI warns ARTEX AI hacking tool is powering global cyberattacks
- MITRE ATT&CK
Frequently asked questions
What is autonomous penetration testing?
It is the use of AI-driven software to plan and perform the steps of a penetration test with little human direction, adapting to what it finds.
How is it different from a vulnerability scanner?
A scanner checks for known flaws and reports them. An autonomous tool tries to exploit weaknesses, chains steps together and decides what to do next based on results.
Is ARTEX a legitimate tool?
ARTEX was published as a penetration testing framework. The FBI says multiple threat actors have used it in attacks since at least August 2026. Whether a tool is legitimate depends on whether its use is authorized.
Can autonomous penetration testing replace human testers?
It can cover breadth and repetition well. Human testers remain better at business-logic flaws, creative attack chains and judging real-world impact. Most teams use both.
How do I defend against AI-driven attacks?
Reduce what exists to be found: fewer exposed assets, no standing privileges, short-lived credentials, and secrets out of code. Then detect machine-speed behavior and automate safe responses.