The rapidly changing landscape of cyber defense and offense is becoming increasingly complex. In recent weeks, the advent of “frontier” AI models has significantly shifted the dynamics of the cybersecurity arms race, leading to a critical need for reassessment of existing security strategies. While these advanced models have the potential to fundamentally transform how vulnerabilities are identified and remedied, they also introduce a daunting “velocity problem,” particularly concerning Operational Technology (OT) and Internet of Things (IoT) environments.
Recent developments have set the pace for this transition, notably the release of two significant AI models: Anthropic’s Claude Mythos Preview and OpenAI’s Daybreak. Launched on April 7, Mythos is accompanied by “Project Glasswing,” an initiative designed to provide early access to defenders in order to stay ahead of the tidal wave of vulnerabilities that this model can uncover. On May 11, OpenAI announced Daybreak, which takes the initiative a step further, not only in discovering vulnerabilities but also in generating patches utilizing the OpenAI Codex engine.
While the effectiveness of these frontier models in identifying vulnerabilities beyond the capacity of human researchers remains debated, there is no doubt that they are capable of discovering vulnerabilities at an unprecedented volume and pace. For operators managing OT systems, the most troubling aspect of AI-generated findings is their ability to work across various operating systems and applications. In contrast to IT systems—which typically use a small number of operating systems—OT and IoT systems are fragmented across over 150,000 distinct operating systems.
OpenAI’s advocacy for a “human-in-the-loop” approach raises important questions about human oversight. Given the staggering speed at which these AI models operate, there is a considerable risk that cybersecurity personnel will be overwhelmed, especially in light of the current shortage of skilled professionals in the field. Additionally, Daybreak’s design to manage tasks beyond human manual capability poses fundamental issues regarding what constitutes effective human oversight in cybersecurity scenarios. The stark contrast between Daybreak’s accessible model and Mythos’s more controlled distribution amplifies safety concerns during this early adoption phase of AI-driven security solutions.
The rapid advancements presented by OpenAI just weeks after the Mythos launch should heighten concern among cybersecurity teams. It raises the question of whether OpenAI’s broader release was motivated by safety considerations or whether it was predominantly a strategic response to keep pace with Anthropic’s marketing moves.
Importantly, these AI models do not inherently alter the nature of cybersecurity; rather, they accelerate its tempo. Organizations have long existed in a state where vulnerabilities are identified faster than they can be patched, and AI serves to exacerbate this challenge. In the realms of OT and IoT, where operational environments are unique and specialized, the potential for chaos increases significantly. While AI-generated patches may be effective for generic software, the custom and intricate nature of OT systems complicates their application.
Considering the scale of the challenge, the fragmented landscape of over 150,000 operating systems necessitates a keen understanding of the intricate interplay between devices and the applications managing them. For tools like Daybreak to prove effective, they must be capable of comprehending this device-to-application integration. Without such contextual awareness, the patches generated on-the-fly risk rendering mission-critical systems inoperable.
Emulating complex environments and deploying rigorous testing is a critical component that current AI models lack. While programs like ARPA-H’s UPGRADE explore automated patch generation within medical systems, they differentiate themselves by emphasizing the development of detailed emulators for pre-deployment patch testing. OpenAI’s Daybreak, however, does not incorporate this essential emulator step. In an OT context, where the deployment of an incorrect patch can result in severe downtime or safety hazards, the absence of laboratory testing before implementing AI-generated solutions poses an insurmountable barrier.
Furthermore, early experiences with AI code generation have revealed a predictable pattern: while these AI systems project confidence, their understanding of complex tasks often falls short. For instance, in personal experiences involving AI for website management, it becomes evident that when faced with obstacles, AI agents tend to fabricating seemingly endless solutions without fundamentally resolving the original issues. This can lead to a confusing array of technical debt rather than creating more secure architectures. With AI-generated code in frontier models, there is a tangible risk of creating complexity that may even overwhelm the original equipment manufacturers.
Looking ahead, it is crucial to recalibrate expectations to align with the current state of these tools. The cybersecurity industry is steadily progressing toward automated vulnerability assessment and risk-based prioritization, with future aspirations of achieving targeted autonomous exposure remediation. While AI has a role to play in this evolution, its immediate deployment should be approached with caution, robust human oversight, and a commitment to laboratory testing environments.
In the meantime, organizations must prioritize multifaceted defensive strategies. Although patching is a vital component of cybersecurity, it should not be viewed as a standalone solution. Foundational controls—ranging from credential management to software updates and non-human identity management—must operate collaboratively to provide compensatory protections when patches fall short. Ultimately, exposure management requires a delicate balancing act between operational necessities, potential opportunities, and inherent cybersecurity risks. While AI serves as a potent tool for increasing the velocity of vulnerability discovery, in the specialized realm of OT, speed without precision is just another vulnerability waiting to be exploited.