After the Old Order Shatters: Security's New Order Under AI and New Opportunities for Practitioners
The cybersecurity industry is now caught in a striking structural paradox.
On one hand, AI is rapidly approaching or even surpassing human performance in traditional security work, including vulnerability discovery, penetration testing, security operations, and reverse engineering. Many security researchers and frontline practitioners are unsure what role they will play in the future. On the other hand, real-world security conditions continue to deteriorate along multiple dimensions. With AI amplifying both sides, offense and defense have become severely asymmetric: no matter how hard defenders try, they cannot stop every attack. The introduction of AI agents is testing the governance capacity of existing permission systems, while credentials and sensitive data are being leaked with increasing frequency.
Why are the “replacement crisis for security roles” and the “systemic failure of organizational defenses” happening at the same time?
The central reason is that the underlying assumptions of the “old order” that sustained the security industry for decades no longer hold. The academic research paradigms, business models, security products, enterprise defense strategies, and divisions of labor built around that stable state are all undergoing a profound reshuffling and reassessment.
When an old equilibrium breaks, the industry inevitably moves toward a new one. This article develops the argument on three levels. First, it begins with frontline security practice to examine how the old order worked, why it failed, and which core foundations and long-term technical directions will define the new order. Second, it introduces an analytical framework from the economics of labor and specialization to explain how AI, as a general-purpose technology, is redrawing task boundaries and how practitioners should shift their roles. Finally, it returns to philosophy and the human scale to ask what individuals should rely on to establish their value and inner steadiness once instrumental rationality is replaced by algorithms at scale.
To understand the present disorder, we must first clarify the logic that sustained the entire security system over the past several decades.
1. The Underlying Logic of the Security Industry: How Did the Old Order Work?
Strip away the dense layers of technical terminology, and the security industry ultimately provides only two fundamental kinds of value: security and trust.
- Security: At its core, security is a contest between attack and defense over a specific objective. Attackers try to break through defenses to achieve a goal, such as stealing data, disrupting system operations, or gaining financially. Defenders work to stop them and preserve the continuity of their assets and business operations.
- Trust: Trust addresses the more fundamental question of certainty: how do we prove that an identity is authentic—that “you are who you say you are,” through authentication and authorization—and how do we prove that data, code, or system state has not been tampered with, through integrity and verifiability? Trust is the foundation on which every security mechanism operates.
These two forms of value do not remain isolated at the technical level. Security conflict and the construction of trust always take place within concrete “organizations,” namely companies and institutions. Around the real needs and conflicts inside and outside these organizations, the security industry has developed a tightly connected value chain with a clear upstream and downstream division of labor over the past several decades.
Link One: Foundational Security and Trust Research (Security Labs and Universities)—Providing Basic Primitives and Technical Foundations
Universities and security labs sit upstream in the value chain. They are not the primary battlefield of real-world attack and defense. Their core role is to explore technical frontiers and provide downstream organizations with foundations for security and trust. To test whether mechanisms are reliable, upstream research typically evolves through a binary interplay between attack and defense, supporting the two core forms of value described above:
- Building the foundations of trust: This includes cryptographic research (cipher design vs. cryptanalysis), protocol security research (protocol design vs. protocol analysis and attacks), and software protection (code obfuscation vs. deobfuscation and reverse engineering). Through continual security validation, this research provides digital systems with reliable foundations for authentication, encrypted communication, and protected execution.
- Advancing security technology: This includes vulnerability research (for example, vulnerability discovery vs. secure design), system security research (exploitation vs. mitigation), malware research (evasion vs. detection), and risk research for emerging technology stacks (such as prompt injection attacks vs. protective mechanisms). This work anticipates the evolution of attack and defense at the mechanism level and builds broadly applicable capabilities for frontline practice.
Link Two: Security and Trust Within Organizations—the Main Battlefield Where Security Delivers Value
Technical capabilities produced upstream can deliver value only when they enter specific organizations. Because high-value data, funds, and business processes are concentrated within organizations, real attack-and-defense conflict inevitably centers on them:
- On one side are goal-oriented attackers: Whether they are financially motivated cybercriminals and ransomware groups or nation-state advanced persistent threat groups with strategic objectives, they act with explicit goals—to obtain sensitive data, hijack business processes, extract illicit gains, or disrupt the availability of critical systems.
- On the other side are defenders operating as a system: In response to external threats, organizations integrate upstream capabilities into a standardized defense model: security governance, compliance, and architecture design (establishing baselines and trust boundaries) → penetration testing and Security Response Center operations (finding and closing gaps) → security operations (threat hunting, incident monitoring, incident response, and vulnerability remediation).
This logic runs through different levels of organizational operations. In enterprise security, the focus is on applying it to enterprise security architecture, intrusion detection, and response. In application security, the focus is on secure application design, vulnerability detection, and remediation. In business security, it centers on telemetry design and the detection and disruption of cybercriminal activity around core business rules.
Whatever the specific context, organizational security has a consistent goal: preserve the trust boundaries of systems and transactions, and raise the attacker's cost of action above the expected return.
Link Three: The Business Logic of Security Vendors (Products and Services)—Standardizing Organizational Defense Needs
Because most institutions cannot independently build a complete security capability, vendors have gradually developed two mature business models around the customer's defense model: product delivery and expert services.
- Security products (governance, protection, and detection): Beyond tools for identity management, encryption, and compliance governance, many protection and detection products are built around a monitoring-and-response loop: collect logs → generate alerts and incidents through rules or threat hunting → have security personnel intervene through incident response by removing attackers and patching vulnerabilities. Most of these products are sold together with security services. Intrusion detection products, for example, are commonly bundled with incident response services.
- Security services (penetration testing and incident response): In penetration testing, security experts simulate attackers to locate weaknesses before a real attack occurs and close gaps. Because of the cost structure and authorization boundaries of commercial services, testers generally use expensive zero-day vulnerabilities cautiously, if they need them at all. When defenses are actually breached, vendors help customers contain the impact and restore systems through incident response services.
Upstream research provides foundational primitives. Customers build systems within their organizations to address real threats. Vendors complete the commercial loop around “detection products driven by log alerts” and “security services that depend on expert experience.” These three interlocking links formed the “old order” that kept the security industry operating steadily for decades.
2. The Failure of the Old Order: Collapsing Assumptions and a Broken Value Chain
This old order rested on three implicit premises: vulnerability discovery required substantial human effort and technical expertise, attack execution had a natural ceiling on efficiency, and the risk and trust boundaries of systems were relatively static.
Advances in frontier models and automation broke all three premises within a short period. The sharp decline in the cost of discovering zero-day vulnerabilities, together with stronger AI-driven automated penetration capabilities, reversed the previous cost structure of attack and defense and directly fractured the entire value chain:
- Lower vulnerability discovery and exploitation costs increase threats, preventing detection-and-response security processes from operating effectively. Large models have dramatically accelerated vulnerability discovery and exploit development, tasks that previously had high barriers to entry. The rate at which unknown risks are exposed is beginning to systematically exceed the rate at which humans can patch them. In the face of frequent, large-scale AI attacks and a continuous flow of zero-day vulnerabilities, the after-the-fact chain of “collect logs → trigger an alert → have a human apply an emergency patch” cannot be sustained because of its inherent response delay. The reality is that as long as defenders remain within the old framework of passive response, even using AI to assist with alert triage will not stop automated attacks from breaking through. A significant share of security vendors, however, depend precisely on detection-and-response products for revenue. When the underlying logic of defense fails, their business models inevitably face severe resistance.
- The new risks introduced by AI are themselves open problems with no clear solutions, creating a generational disconnect in existing defenses. Consider the abuse of dynamic permissions after enterprises introduce AI agents, the uncontrolled spread of credentials, blurred data security boundaries, and the failure of established behavioral baselines. Traditional perimeter firewalls, web application firewalls, and static rule engines operate at the wrong level of abstraction for these new systems.
- Human-delivered penetration testing services face a reassessment of their value to customers. On the surface, the growth of system risk appears likely to increase demand for testing services. A closer look reveals the danger. On one hand, traditional point-in-time manual penetration tests cannot exhaustively identify or resolve systemic risks in new architectures. On the other, large models are rapidly flattening the technical barriers to both vulnerability discovery and penetration testing. Tasks that once required many hours from senior experts can now, to a large extent, be completed end to end by models from high-level intent expressed in natural language. When customers can use models directly to perform routine tests at extremely low marginal cost, a penetration testing business model based on selling expert hours faces a severe challenge to its core value.
3. The Foundations of a New Order: Security by Default, Defenders First, and Core Research Directions
When “monitor and remediate after the fact” loses its effectiveness and “point-in-time manual penetration testing” can no longer form a durable barrier, what foundations should support a new order for the security industry?
The new order depends on two fundamental paradigm shifts: Security by Default, meaning default security and intrinsic trust, and Defenders First, meaning proactively discovering risks and moving remediation forward.
- Security by Default (default security and intrinsic trust): The old path—allowing systems to launch with latent weaknesses and relying on runtime monitoring and patches afterward—is no longer sustainable. Security and trust must therefore be established as default properties from the beginning of system architecture, communication protocol, and permission model design. Deterministic security architectures, least-privilege constraints, and microsegmentation can reduce an enterprise's exposure and the blast radius of a threat.
- Defenders First (completing the scanning and remediation loop before attackers act): If AI has dramatically lowered the execution barriers to vulnerability discovery and penetration testing, defenders can no longer wait passively for alerts. They must act before attackers do, using AI to continuously and comprehensively conduct automated vulnerability scanning, attack simulation, and rapid remediation across their own codebases, infrastructure, and business flows. Only by finding and eliminating weaknesses before attackers can defenders regain the initiative, sharply reduce the number of risks exposed after launch, and keep detection-and-response security operations viable.
Around these two foundations, the security problems that offer genuine long-term defensibility and deserve sustained work and research fall mainly into three directions.
Direction One: Help Build the New Order
Since existing frameworks are already out of step, the most urgent task is to establish a new order within organizations for the AI era. This includes solving open problems such as building entirely new governance systems for agents, truly embedding Security by Default into enterprise security and business architecture, and implementing a Defenders First strategy within organizations. The main difficulty is not the technology itself. It is working like a forward-deployed engineer to put the new order into practice inside organizations full of conflicts and competing pressures. People doing this work need both strong professional expertise and exceptional coordination skills. They must be able to balance the interests of different parties and make things happen. This person may be the head of security at the customer organization or someone at a security vendor playing a role similar to an FDE at an AI company.
Direction Two: Strengthen AI for Security
Using AI to implement proactive scanning and remediation and a new generation of security operations has very different lifecycles and competitive barriers at three levels:
- Enterprise systems implementation (rebuilding AI-native SOCs, SDLCs, and red teams): Make the Defenders First loop work in real enterprise production environments, deeply embedding automated scanning, risk simulation, and automated remediation into the software development lifecycle (SDLC) and the security operations center (SOC). This level is tightly coupled with specific business and engineering systems, creating an extremely high implementation barrier.
- Peripheral tools and middleware (building vulnerability discovery or penetration testing harnesses and application wrappers): Orchestrating external workflows around narrow scenarios such as vulnerability discovery and penetration testing is, in essence, the construction of general-purpose peripheral tools. I am cautious about the long-term value of this direction. As foundation models continue to advance, the lifecycle of these capability-enhancing peripheral frameworks, or harnesses, designed to compensate for current model limitations is often very short. Capabilities filled in today through complex engineering orchestration are often absorbed natively in the model's next iteration. As models become increasingly able to respond directly to high-level intent, the independent room for pure tool-wrapping layers will continue to shrink.
- Foundation model layer: This is also one of my current research areas: directly strengthening the model's native reasoning and offensive and defensive capabilities for AI for Security. Compared with wrapping external workflows, work at the model layer is more fundamental. In my view, this kind of work is best led by researchers who have participated in real frontline attack-and-defense work and understand the essence of security.
Direction Three: Strengthen the Security and Controllability of AI Systems
As foundation models and agents continue to grow more capable, the new order must confront a central question: how do we address new security threats such as prompt injection, ensure that AI systems remain trustworthy and under control, and even prevent highly capable systems from escaping control and harming human society? This direction also has three levels:
- Application governance layer: Constrain the scope of agents' permissions and their boundaries for accessing sensitive data within enterprises, ultimately bringing this work together with the next generation of organization-wide security governance.
- External mechanism layer: Design external sandboxes and access-control mechanisms that strike a balance between security and efficiency.
- Foundation model layer: Address security and controllability directly within the model. Its technical stack closely overlaps with the model layer of AI for Security, with the central focus on safety alignment to strengthen the model's intrinsic security.