The Industrialized Attacker: Why Cyber Defense Must Be Retooled for Speed

Authors: Scott Harrell, President and CEO, Infoblox, and Phil Venables, Board Member, Infoblox and Partner, Ballistic Ventures

The word “frontier” has taken on new meaning in the age of AI. Every new release expands what is possible with foundational models, making the extraordinary become ordinary. We’re living through a disruption that will potentially be akin to the rise of industrial age; but disruption is often messy and while much new opportunity will be created, there are also new risks created. Nowhere is this truer than in cybersecurity. AI both requires and allows us to fundamentally rethink cyber.

While there is appropriately much fear and concern about the power of frontier models in the hands of bad actors, there is also a lot of good to already be celebrated and looked forward to. Legacy code bases that have long been the source of vulnerabilities and exploits are being completely rewritten in modern, less exploitable languages. Agents are being applied to cyber operations to help mitigate what has been a chronic shortage of cyber talent. Patching is being automated at a scale and speed that no one thought was possible. Software development lifecycles are becoming faster and more secure through the application of agentic code reviews. Startups and established security vendors across the world are leveraging these faster software lifecycles to unleash a huge wave of innovation aimed at creating new approaches to protect AI systems and defend against AI-driven attackers. These are just some of the reasons we are long-term optimists about the future of cybersecurity.

However, in the shorter term it is a race against attackers to figure this out. To win this race, defenders must anticipate what is next and look for early clues as to how attackers are leveraging AI and where legacy defenses are likely to break or become increasingly ineffective. To deal with the speed of change needed, security organizations will have to be progressively more agile and willing to challenge and disrupt the status quo. This means rapid evolution of security organizations’ internal operations and processes, organizational paradigms and their tools and vendor ecosystems. In the past, this type of broad scale, fundamental change has only happened after something in security had already become an epidemic, or regulations forced the change. This urgent need for change across so many dimensions in parallel is part of the reason we are short-term pessimists.

The Forces Reshaping Cybersecurity Now

AI is driving converging forces that security leaders now must manage in parallel.

The tidal wave of vulnerabilities plays a large role: models, including open-weight systems accessible to anyone, can already find weaknesses at scale in even the most scrutinized codebases. In July 2026, OpenAI disclosed that a combination of its models escaped an isolated testing environment by exploiting a zero-day in OpenAI’s own application stack, then reached Hugging Face’s platform while goal-seeking an answer to an evaluation online, a behavior it labeled “reward hacking.” OpenAI’s application stack is about as scrutinized as software gets, and its own models still found a hole in it without being asked to; it’s a preview of how routine zero-day discovery is becoming.

We’re also seeing the industrialization of multi-stage campaigns. Adversaries are using AI to chain together exploits and pursue targets at a scale no human operator could sustain. Very soon, we’ll stop labeling incidents as “AI-driven attacks” because that will simply describe all serious campaigns. Sysdig’s Threat Research team documented JADEPUFFER in July 2026, the first known fully autonomous, end-to-end AI-driven ransomware operation, in which an LLM agent ran a destructive extortion playbook with no human directing the individual steps.

Every vulnerability is a viable target of opportunity, and the time between disclosure and exploitation is compressing toward zero. The old playbook of layered defense, patch cadence and herd immunity was built for a resource-constrained adversary. That adversary no longer exists.

The Missing Piece: Anticipating What’s Next

Most of the conversation around defending against AI-driven attackers focuses on using AI to accelerate detection and response. This is true of vulnerabilities and patching, breaches and remediation, novel malware and blocking of their new indicators of compromise. These are critical actions to take against today’s threats, but the speed of AI adoption and its evolution requires we anticipate what is beyond these threats that are already present and understand what could be needed to deal with what’s ahead.

First, we are sliding into a world of fake everything: fake companies, fake workers, fake content and deepfakes indistinguishable from real interactions. Over the next 12 to 18 months, as the industry makes genuine progress closing exploitable technical vulnerabilities, adversaries will pivot. We may shrink the exploitable software attack surface, but the human attack surface is not going anywhere.

AI makes that pivot terrifyingly scalable: hyper-personalized social engineering at industrial volume, aimed not just at people’s work identities but at their personal ones. In August 2026 alone, Google researchers tracked a single extortion crew, operating under the umbrella UNC6671, that phoned employees at more than 200 organizations on their personal mobiles over five weeks, posing as colleagues or IT staff and steering them to spoofed login pages built to harvest credentials and multi-factor authentication (MFA) codes.

The employee protected by every corporate control can still be reached through a deepfake voice call to their personal phone, or a synthetic relationship built on scraped public data. This isn’t hypothetical. As GetReal Security has documented, a single deepfaked video call has already triggered a $25 million wire transfer, and a fabricated audio clip disrupted a national election in Slovakia. When the attack surface is what people believe, across every screen they touch, there is no patch that closes the gap.

Second, attackers have always been more resource-constrained than people realize, but that constraint is now gone. A ransomware gang that could exploit ten organizations a week can now run AI agents and target ten thousand. Attacker agents are already available for purchase, but also are increasingly available freely in open-source tool repositories, according to recent reporting from ESET. Open-weight code assist tools will allow for an infinite creation of variants of these agents and their associated behaviors, as well as custom code that can be deployed at scale. Additionally, “double” agents will be created to imitate legitimate users or corporate-sanctioned agents for specific targets, leveraging socially engineered credentials to bypass internal controls.

This isn’t theoretical: in July 2026, Zenity Labs disclosed “AgentForger,” a flaw letting one phishing link forge an agent inside ChatGPT Enterprise that silently inherited a victim’s authorized connectors and then impersonated them internally, with no new credentials or alerts triggered.

Third, the time from intrusion to exfiltration, ransom or other types of attacker-driven disruption will continue to collapse, dropping from roughly an hour to mere minutes or even seconds. CrowdStrike’s 2026 Global Threat Report puts a number on that trend: average e-crime breakout time fell to 29 minutes in 2025, a 65 percent jump in speed year over year, with the fastest breakout on record clocking in at 27 seconds and one intrusion reaching exfiltration in just four minutes. That acceleration isn’t a one-off: AI-enabled adversary operations rose 89 percent year-over-year, with actors like e-crime group PUNK SPIDER already using AI-generated scripts to accelerate credential dumping and erase forensic evidence before defenders finish triaging the alert.

None of this bodes well for the efficacy of most legacy defenses. Applying AI to existing tools is a good first step, but does little to stem the tide of emerging threats posed by AI-enabled attackers. Detect and respond methodologies will be too slow and prone to error, even when leveraging AI. The herd immunity model most security tools rely on for protection will essentially disappear once every attack is leveraging a zero day and mass customization of malware and rogue agents are the norm.

Security is unique in that you must always deal with an active adversary; AI enables the adversary to produce and act with endless variety. Speed must be improved, but this alone will be insufficient. Defenders must seek new methods or double down on existing ones that anticipate this variety and proactively counter it.

What’s Next: An AI Resilient Defense Strategy

The baseline must be complete, not aspirational. Authentication, segmentation and application security have been on every framework for twenty years. But an industrialized attacker will find the three percent of endpoints you skipped, the one segment you left open or the account without MFA. “Reasonably good” is no longer a defensible standard.

Preemptive security must become a priority. We have historically focused most of our energy and investment as a security community on how we detect and respond to threats.  In the age of AI, this will be too late. Attackers will continue to automate to recursively compress the time from initial intrusion to action. We need to ask how we can disrupt and inhibit an adversary before they act.

Every attack requires preparation, even in the age of AI. Command-and-control domains get registered, brand impersonation in social media gets created and phishing scaffolding gets stood up days or weeks in advance, which leaves traces. Before this malicious infrastructure can be weaponized we must seek to disrupt it.

Move upstream and stay there. DNS-layer visibility combined with AI-driven analysis of infrastructure behavior creates the ability to identify and block adversary tooling before it fires, including the infrastructure behind social engineering campaigns targeting your people’s personal and professional identities. Each hour of lead time is leverage.

If you can identify and block the domains behind a synthetic persona campaign before it reaches your people at work or at home, you have disrupted the operation at its root rather than chasing its consequences. The cybersecurity industry has spent decades building better locks; what we’ve underinvested in is tearing down the scaffolding before anyone climbs it.

Embrace deception and moving-target defense. AI-driven attackers are more susceptible to deception technology than human attackers are. They follow patterns. They get confused by dynamic environments. Honeypots and moving infrastructure are showing genuine results against automated adversaries and deserve serious reinvestment. We’re already seeing this pay off: AI-generated decoy data is misleading automated ransomware operators long enough for response teams to isolate and restore the real environment untouched.

Govern the agentic layer now. Agent identity, access controls and behavioral monitoring are not future problems. Workload identities—applications, scripts and agents—already outnumber human identities by a large margin, and most are still secured with static, hard-coded credentials rather than anything resembling MFA. Every workflow handed to an AI agent without a clear governance model is a potential pivot point for an adversary already probing those systems.

Each agent that belongs to and matters to a business should have a name, just as each user does. That name is the basis of identity, with authentication built on top of it. The infrastructure, and DNS in particular, is a natural place for that naming to happen. Efforts like DNS-AID anchor an agent’s name to a domain the organization already controls and use DNSSEC to cryptographically verify the record, so any action an agent takes can be traced back to an accountable owner. That same naming capability can double as a kill switch: if an agent becomes jailbroken or otherwise goes rogue, blocking resolution to its domain cuts it off before it can act further.

The right model isn’t logging what an agent already did. It’s an enterprise agentic control plane that intercepts what an agent is about to do, watches for drift between what it was authorized to do and what it’s now doing, and revokes access the moment that drift appears. The Cloud Security Alliance’s new AARM framework for securing the agentic control plane is a good signal of how fast this new thinking is moving from theory to standard.

Treat speed as an organizational discipline. The real bottleneck is almost always slow approval chains, siloed teams and governance built for a different era. Speed without adaptability isn’t a strategy. Static playbooks, however fast, can’t match adversaries generating endless variety of attack paths; only variety can absorb variety.

But speed also has to run on good data, or it becomes a liability instead of an advantage. Automating on top of bad data compounds the chaos and any mistakes exponentially. Most configuration management databases (CMBDs) have long averaged roughly under 70 percent data accuracy. That’s no longer acceptable once AI is making the decisions: it needs data that is as close to 100 percent accurate and real-time as possible, or it risks making the wrong call at scale.

The same is true of the context feeding security data lakes, security information and event management (SIEM) and similar systems. Gaps there get compounded too, and the effects can be catastrophic. They also erode trust in the automation itself, forcing the kind of constant human intervention that can’t scale against AI-driven adversaries.

The Questions Leaders Should Be Asking

The fundamental question every CEO and board member should ask: “Are we still operating on the assumption of a resource-constrained adversary?”

Four questions to pressure-test where you stand:

• Is your security baseline 100 percent complete, not just reasonably good?

• Do you have the capability to detect and disrupt adversary infrastructure before it is used against you, including infrastructure behind AI-generated social engineering campaigns?

• Are you prepared for adversaries who will target your people’s personal digital identities, not just their corporate ones?

• Have you retooled your organization, processes and governance for the speed this environment demands?

The window to act is narrowing, but it is still open.