COMMENTARY: I often view security as a race. Cybercriminals innovate attacks; defenders innovate security frameworks to stop them. Both try to stay one step ahead.Now, there’s a new entrant in this arms race. I’m referring to frontier AI models acting as agentic attackers. They are fast, relentless, and capable of running multistage campaigns against enterprises at a scale individual bad actors cannot match. Defenders must move at the same speed. That’s the new benchmark.[SC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Read more Perspectives here.]Most industry conversations about AI security center on securing the AI model. This matters, but it misses a bigger shift, and that’s the industrialization of attacks.What’s changing?Back in the day, advanced attackers needed specialized skills to find a vulnerability, check whether they could exploit it, and turn it into a workable exploit. This skill was not easy to find and created a barrier between low- and advanced-level attackers.Frontier models have all but eliminated that barrier. Our threat intelligence team calls this new generation “the zero-knowledge threat actor.” These are malicious actors who one year ago may not have been able to run a sophisticated attack, but can now do so with a prompt. AI has not only helped attackers move faster; it helps them deploy techniques they could not previously use, at far greater scale.Here's what we're seeing:
While AI makes the attackers' job easier, we should not see it as having unlimited leverage. AI models require time and money, which means adversaries still have to prioritize. But in the larger scheme, it’s a small constraint compared to the advantage AI gives attackers.Why single-purpose tools struggleHere’s where the traditional security architecture runs into trouble.It’s not necessarily the individual tools: it’s how they are stitched together and their inability to operate as a system.In my experience, the typical CISO manages numerous point products. Each product has its own console and data set, creating only a partial view of the environment. This fragmentation was costly even before the arrival of agentic threats. Against an attacker that moves in real-time and chains together reconnaissance, exploitation, and lateral movement autonomously, fragmented stacks are often catastrophic. By the time the various tools report their piece of the picture, the attacker may have already moved on.We can no longer rely on static signatures. Defense must focus on behavior and intent because an attacker's model can rewrite its signature during an attack.What defenders must doToday, it’s not about investing in a new product category, but about building a different architecture, which I encourage organizations to evaluate regardless of vendor:
Organizations running legacy infrastructure that’s difficult to patch are more exposed. Attackers are giving agents responsibility for entire attack chains, including reconnaissance, exploitation, and credential guessing, enabling them to attack hundreds of targets at once across a range of organizational systems. In this scenario, defenders have to move from detection to protection in near real time, across every system.Every major shift in technology has forced us to rethink how we manage risk than others. Cloud changed how we think about identity. Mobile changed how we think about endpoints. In my view, agentic AI will change the architecture question: it’s not just about what we defend, but how quickly we can defend it.I do not believe the organizations that come out ahead will use the most AI. They will rebuild their security architecture around real-time visibility, behavioral detection, and automated response. So when the next zero-day gets discovered by a model rather than a researcher, the response cannot depend on a human noticing it first.Etay Maor, vice president of threat intelligence, Cato NetworksSC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Each contribution has a goal of bringing a unique voice to important cybersecurity topics. Content strives to be of the highest quality, objective and non-commercial.
- Finding and linking vulnerabilities into a complete attack path can now take a fraction of the time it once did.
- AI now helps attackers generate malware and phishing tools faster and continuously mutate their code. Instead of a fixed signature a security tool can recognize, each version looks a little different and can evade detection.
- The time between discovering a flaw and turning it into a working exploit has shrunk to zero. It's undermining the disclosure-and-patch cycle most organizations depend on to stay ahead of attackers.
- In one reported case, Google identified malware dubbed PROMPTFLUX, that queried an AI model during mid-attack to rewrite its code and evade detection.
- A unified view of telemetry: So the team can interpret signals from network, endpoint, identity, and application in one place.
- The ability to patch and protect within hours of disclosure: A major shift from the week-long schedule that manual, appliance-by-appliance rollouts can require.
- Behavior-based detection: Reduces dependence on fixed signatures.
- Automated, continuous response: So the team can keep pace with machine-speed attacks.
- Streamlined defense: Teams need a defense that does not require more analysts whenever attack volumes increase, because agentic attacks are designed to generate volume at scale.
