The Industry Is Racing Toward Automation. We're Drawing a Line.
A warning about autonomous AI penetration testing, the growing pressure to replace human expertise, and the consequences of allowing offensive technology to make decisions inside critical systems.
Never the Tester.
Penetration testing should be conducted by experienced security professionals who understand the environment, control exploitation decisions, and validate actual exposure. Technology can support their work. It should never replace the judgment required to perform it responsibly.
I've spent nearly a decade building Redbot Security around a simple belief. Penetration testing should be performed by experienced security engineers who understand what they're testing, why they're testing it, and the potential consequences of their actions.
Today, I'm watching the cybersecurity industry move in a direction that genuinely concerns me. We've seen experienced security professionals lose their jobs, budgets tighten, and a growing number of penetration testing providers enter an increasingly crowded market. Now autonomous artificial intelligence is becoming the latest selling point, with vendors promising faster results, continuous testing, and lower costs than traditional human assessments.
At Redbot, we're beginning to see the impact directly. Some customers are exploring autonomous AI penetration testing as a lower cost alternative to manual assessments, despite concerns we've raised about the potential risks of autonomous exploitation. I understand the financial pressure these organizations face. Every business wants to operate more efficiently, and artificial intelligence is an extraordinary technology. Its capabilities are expanding rapidly, including in offensive security.
But there's a fundamental difference between using technology to make an experienced engineer more effective and allowing autonomous software to decide how to attack an organization's systems. That distinction matters, particularly when we're talking about production environments, sensitive information, and critical infrastructure.
Just because an autonomous system can find a way into your infrastructure doesn't mean it should be trusted to decide what happens next.
Recent findings from the FBI and Google have made this discussion even more urgent. Their research documents how powerful autonomous offensive technology has become and how threat actors are already adapting these capabilities for real cyberattacks. I believe those findings should prompt a serious conversation among CISOs, business leaders, security providers, and anyone responsible for protecting critical systems.
The FBI Has Documented the Weaponization of Autonomous Penetration Testing
On November 13, 2026, the Federal Bureau of Investigation issued a FLASH advisory concerning ARTEX, an autonomous penetration testing framework being weaponized by malicious actors. [1] The advisory describes how offensive automation can support multiple stages of cyber intrusions, accelerating activity that traditionally required direct human involvement.
The FBI's findings address malicious activity affecting organizations across several industries, including financial services and other sensitive sectors. The significance extends beyond the incidents themselves. Autonomous offensive tools can reduce the amount of time and manual effort required to identify weaknesses, develop attack paths, and pursue access to additional systems.
The cybersecurity community has spent years warning organizations about how attackers can combine individual vulnerabilities into larger compromises. Autonomous systems introduce the potential to execute portions of those sequences much faster.
FBI FLASH Advisory, November 13, 2026
Threat Overview of ARTEX Autonomous Penetration Testing Framework.
It's important to be precise about what this warning establishes. The FBI is documenting criminal weaponization of offensive technology. It is not declaring that every legitimate commercial autonomous penetration testing platform is inherently unsafe.
Nevertheless, the findings demonstrate the consequences of giving increasingly capable software offensive functionality. Those capabilities should be carefully understood before they are authorized inside an organization's infrastructure.
Google's Findings Reinforce the Growing Concern
On November 16, 2026, Google Threat Intelligence Group published research describing a significant shift in the operational use of agentic AI for malicious cyber activity. [2] The research documents how attackers are beginning to use autonomous workflows for activities extending beyond simple code generation or assistance with individual tasks.
Google and Mandiant have also examined offensive frameworks including ARTEX, CAIRN, CyberStrikeAI, and PentAGI. [3] These technologies demonstrate how software can coordinate reconnaissance, vulnerability research, exploitation, and other stages of an intrusion.
This changes the speed and scale at which certain offensive activities can be performed. It also raises important questions about authorization, oversight, containment, and the operational consequences of actions taken by autonomous agents.
The ability to execute an exploit is not the same as the judgment required to determine whether it should be executed.
I want to emphasize that these findings concern malicious activity. They do not establish that legitimate vendors using autonomous technology lack safeguards or are responsible for the attacks described.
What they do establish is the increasing capability of autonomous offensive systems. That alone should be enough to encourage organizations to examine how such systems are controlled before allowing them to operate against sensitive environments.
Google Threat Intelligence Group
We're Losing Experienced People While Selling the Idea That We Need Fewer of Them
One of the developments that concerns me most is the growing pressure to reduce human involvement in cybersecurity at the same time our industry is facing workforce reductions and budget constraints. The 2025 ISC2 Cybersecurity Workforce Study documented ongoing organizational layoffs and security budget reductions. [4] These pressures are real, and they influence how companies purchase and deliver security services.
Meanwhile, new penetration testing providers continue to enter the market, some offering increasingly automated assessments at prices that traditional consulting firms cannot easily match. Competition and innovation are healthy, but the barrier to marketing a penetration testing service is not the same as the barrier to delivering one responsibly.
An attractive website, an automated report, or a promise to complete an assessment in hours does not establish the same level of technical depth, operational maturity, or professional judgment that a customer should expect from an experienced offensive security provider. Those qualifications have to be demonstrated through methodology, engineering expertise, testing controls, and verifiable results.
I have no problem with new companies entering this industry. Some are doing genuinely impressive work. What concerns me is the possibility that penetration testing becomes increasingly commoditized, with customers encouraged to compare services primarily on price and speed without fully understanding the differences in how assessments are actually performed.
The cheapest penetration test is not necessarily the safest or most effective assessment.
Organizations should understand who is conducting the test, how exploitation is controlled, whether findings are independently validated, and what safeguards protect production systems before comparing price alone.
The Companies Moving Penetration Testing Toward AI
This shift is not limited to small startups. Established cybersecurity companies and newer technology providers are investing heavily in AI powered testing, autonomous exploitation, and automated security validation.
Their products are not identical, and the level of human involvement varies significantly. Some provide autonomous exploitation capabilities, others combine AI with human validation, and others primarily use automation to support experienced security professionals. These differences are precisely what customers should understand.
| Company | Publicly Marketed Approach | Source |
|---|---|---|
| Aikido Security | Markets Aikido Attack as an autonomous AI penetration testing platform focused on accelerated assessments, exploitation validation, and continuous testing. The company emphasizes speed and lower costs compared with traditional penetration testing, while also describing testing safeguards and controls. | Aikido Attack |
| XBOW | Markets autonomous offensive security agents capable of discovering, testing, and validating vulnerabilities in applications and APIs. The company emphasizes autonomous execution, continuous coverage, and verified findings. | XBOW |
| Horizon3.ai | Offers NodeZero, an autonomous penetration testing platform designed to identify and validate exploitable attack paths across enterprise environments. The company markets safeguards for production testing. | NodeZero |
| NetSPI | Promotes human led, AI accelerated penetration testing and continuous security validation. Its model combines experienced testers with AI driven testing capabilities and automation. | NetSPI |
| Synack | Offers Sara AI Pentesting, combining agentic testing capabilities with human researcher expertise and validation. Synack and NetSPI announced a combination in 2026. | Synack |
| Rapid7 | Invests in AI and automation across its broader cybersecurity portfolio. Its traditional penetration testing services continue to emphasize experienced human consultants. Rapid7 should not be characterized as a fully autonomous pentesting provider. | Rapid7 Services |
| Pentera | Provides automated security validation designed to identify and demonstrate exploitable exposure across enterprise infrastructure. Its approach focuses on adversarial validation through automation. | Pentera |
| Escape | Markets AI powered application and API security testing, including agentic capabilities and continuous validation. Its focus includes identifying application security weaknesses and business logic risks. | Escape |
| Terra Security | Promotes agentic penetration testing with human involvement, emphasizing recurring testing and a combination of AI capabilities with security expertise. | Terra Security |
Editorial clarification: This table summarizes public product positioning and does not represent an independent safety evaluation of the listed companies. Inclusion does not suggest malicious conduct, inadequate safeguards, or involvement in the criminal activity documented by the FBI and Google. The companies listed offer different testing methodologies and levels of human oversight. Customers should evaluate each offering individually.
What Does Human Led Really Mean Anymore?
As more penetration testing companies introduce AI into their services, I've noticed something interesting. The phrase "human led" is becoming increasingly common. It's reassuring language, but I believe organizations deserve to understand exactly what it means.
There is a significant difference between an experienced security engineer who manually conducts an assessment, an engineer supervising automated exploitation, and someone reviewing findings after an autonomous platform has completed its work. All three approaches can involve a human, but they are fundamentally different testing models. The distinction becomes particularly important when testing production infrastructure, sensitive applications, or environments where an unexpected action could disrupt operations.
At Redbot Security, our engineers make the decisions throughout the assessment. They understand the authorized scope, evaluate potential consequences, and determine how far an exploitation attempt should proceed. They also understand that confirming a vulnerability does not always require demonstrating every possible consequence of exploiting it. Sometimes the responsible decision is to stop, document the evidence, and work with the customer to determine whether additional validation is appropriate.
A human reviewing the results afterward is not the same as a human controlling the risk while testing is happening.
I'm not suggesting that every autonomous testing platform lacks safeguards or that automation cannot improve security outcomes. It certainly can. But the industry needs greater transparency about what is actually being performed by an engineer, what is being performed by software, and who has authority over the actions taken during an assessment.
Critical Environments Are Not Testing Grounds for Uncontrolled Automation
Consider what happens when an autonomous system identifies a viable exploitation path inside a production environment. From a purely technical perspective, the next step may seem obvious. The system has identified a weakness, developed an attack sequence, and determined that executing another action will produce additional evidence. But offensive security is not simply a process of following every technically available path.
A vulnerable server may support a hospital application. A network service may connect to industrial equipment. An identity platform may control access for thousands of employees, while a seemingly ordinary application dependency may support critical business operations. Understanding these relationships is essential before attempting exploitation. A technically successful attack can still be an operational failure if the assessment causes damage that could have been avoided.
This concern is particularly relevant to operational technology and industrial control systems. NIST SP 800-82 Revision 3 emphasizes that these environments have unique safety, availability, and reliability requirements that must be considered when applying cybersecurity practices. [5] Offensive testing in these environments requires careful planning, clearly defined boundaries, and an understanding of the systems that could be affected.
Experienced security engineers recognize that the goal is not to exploit everything simply because an opportunity exists. The goal is to establish meaningful evidence of exposure while respecting the operational constraints of the environment. That judgment, developed through experience and direct accountability, is central to how Redbot approaches penetration testing.
Authorization to test should never be confused with authorization to introduce unnecessary risk.
Organizations considering autonomous offensive testing should independently assess scope enforcement, production safeguards, credential handling, sensitive data access, emergency termination capabilities, and the level of human control maintained throughout an engagement.
We Didn't Build Redbot Around Automation. We Built It Around Experience.
When we founded Redbot Security in 2016, the goal wasn't to build another vulnerability scanning company or produce automated reports that customers would struggle to interpret. We wanted to create a security firm where experienced engineers could perform meaningful manual testing, validate actual attack paths, and provide customers with clear evidence of their exposure.
That approach has shaped the company for nearly a decade. We've spent years refining our scoping process, building our engineering team, improving reporting standards, and developing the discipline required to work responsibly inside customer environments. Our services extend across enterprise networks, applications, APIs, cloud infrastructure, identity platforms, AI systems, and operational technology. Each environment requires an understanding of its own architecture, dependencies, and risks.
Our penetration testing methodology is centered on 100% human controlled manual exploitation and validation. Our engineers use supporting security tools where appropriate, but we do not delegate consequential exploitation decisions to autonomous AI agents. This is an important distinction, and one we intend to continue emphasizing as the market evolves.
We are not opposed to innovation. Artificial intelligence has enormous potential to improve cybersecurity, and Redbot actively researches AI threats, vulnerabilities, and connected systems. But adopting new technology should not require abandoning the professional judgment that responsible offensive security depends on.
Finding Vulnerabilities Is Not the Same as Validating Exposure.
That statement has been central to Redbot's philosophy for years. We're not adopting it because AI penetration testing has become popular. It reflects what we believe customers should expect from a serious offensive security engagement.
The Lowest Price May Become the Most Expensive Decision
I understand why companies are exploring alternatives to traditional penetration testing. Budgets are under pressure, security departments are expected to accomplish more with fewer resources, and the promise of faster assessments at a fraction of the cost is attractive. No responsible business leader should ignore opportunities to improve efficiency.
But security testing is not a commodity where the lowest price automatically represents the best value. An organization purchasing a penetration test is placing a tremendous amount of trust in the provider. It may be granting access to sensitive information, exposing production systems to exploitation attempts, and authorizing activity that would otherwise be considered hostile. Those responsibilities should never be taken lightly.
Before purchasing autonomous AI penetration testing, organizations should understand who controls the assessment, how exploitation decisions are authorized, what technical safeguards protect sensitive systems, where assessment data is processed, and who assumes responsibility if something goes wrong. These questions should be answered with evidence, not simply assurances in a sales presentation.
There may be situations where autonomous testing is appropriate and provides valuable additional coverage. But it should be selected because the organization understands the technology, its limitations, and the safeguards in place, not simply because a lower price makes the decision easier.
Saving money on a penetration test means very little if the testing itself introduces risks your organization never understood.
The Industry Is Changing. Our Standards Are Not.
After nearly ten years building Redbot Security, I have watched this industry evolve through countless technology shifts, new competitors, changing customer expectations, and significant changes in how security services are delivered. Innovation is necessary. Competition is healthy. And technology that genuinely improves security outcomes deserves serious consideration.
But I also believe the cybersecurity industry has a responsibility to challenge developments that could introduce unnecessary risks to customers. As offensive technology becomes more autonomous, transparency, professional judgment, and accountability become more important, not less.
Redbot will continue investing in experienced engineers, responsible testing methodologies, and research into emerging threats. We will continue evaluating new technology without allowing market pressure to dictate our standards. And we will continue advocating for clear distinctions between automated vulnerability discovery, autonomous exploitation, and the manual validation of real business exposure.
To organizations responsible for protecting critical infrastructure, financial services, sensitive information, and essential business operations, my message is simple. Know who is testing your systems. Understand what they are authorized to do. Demand evidence that their methods are appropriate for your environment. And never confuse the ability to exploit a vulnerability with the judgment required to do it responsibly.
At Redbot Security, we have made our position clear.
AI Is the Target.
Never the Tester.
Senior engineer led penetration testing. Human controlled exploitation. Evidence backed security validation. Accountability at every stage.
Critical environments deserve experienced judgment, not uncontrolled experimentation.
Founder & CEO
Redbot Security
Sources & References
This statement draws upon government cybersecurity advisories, published threat intelligence, industry research, and product information from the companies discussed. Research findings, vendor claims, and Redbot Security's editorial opinions are distinguished throughout the article.
Published November 18, 2026. Vendor descriptions refer to publicly marketed capabilities and should not be interpreted as independent assessments of product safety. This article represents the views of Redbot Security's Founder & CEO on responsible offensive security testing and the importance of experienced human judgment.

