The cybersecurity industry is approaching an uncomfortable inflection point. AI is rapidly becoming capable of doing work that once required highly skilled security researchers: navigating unfamiliar codebases, finding vulnerabilities, building exploit chains, testing hypotheses and sustaining complex operations for long periods of time. Frontier AI models are not just about discovering vulnerabilities anymore. They can autonomously carry out a whole range of advanced cyber tasks.
OpenAI recently described the situation succinctly: the window defenders have to prepare for AI-driven attacks is narrowing. Its expanded Daybreak program and GPT-5.6-Cyber are designed to put increasingly capable cybersecurity AI into the hands of trusted defenders before offensive use of comparable technology becomes widespread.
At almost exactly the same time, another debate is intensifying around open-weight AI. Meta, Nvidia, Microsoft and other technology companies have argued that policymakers should avoid premature restrictions on open-weight models, while concerns persist that powerful open-weight systems could make sophisticated capabilities, including cyber capabilities, more accessible. The issue has also become entangled with geopolitical competition. Chinese open-weight models are increasingly competitive with American systems, raising the question of whether restricting American open models might actually weaken the U.S. AI ecosystem rather than make it safer.
It is an important debate. But from a cybersecurity perspective, it risks becoming the wrong debate.
The central question should not simply be whether powerful models should be open or not, or whether cybersecurity capabilities should be made available to researchers. The more consequential question is this:
What happens when machines become better attackers than humans, and can operate at machine speed?
Because that is the security problem enterprises ultimately have to solve.
When the Attacker Is No Longer Human Speed
OpenAI’s Daybreak announcement provides a glimpse of where this is heading. GPT-5.6-Cyber is designed for sophisticated security research, including vulnerability discovery and exploit development. OpenAI reports that the model can perform complex security tasks across large codebases and has already helped researchers identify previously unknown vulnerabilities in real-world software. Its evaluations also measure capabilities such as authentication bypass, privilege escalation and exploit-chain development.
These capabilities can be extraordinarily valuable to defenders. AI could dramatically shorten the time required to discover vulnerabilities, understand their impact and produce patches.
But the same technological trajectory changes the economics of attack.
Traditional cybersecurity has implicitly benefited from a limitation we rarely discuss: attackers have historically been constrained by human bandwidth. Reconnaissance takes time. Testing credentials takes time. Understanding an unfamiliar environment takes time. Finding paths between systems takes time. Developing an exploit takes time.
AI begins to remove those constraints. Imagine autonomous systems continuously examining thousands of targets, testing assumptions, adapting when techniques fail, analyzing configurations, correlating leaked credentials, identifying privilege-escalation paths and trying alternative approaches without getting tired, distracted or discouraged.
The question then becomes less about whether an AI system can perform one spectacular hack and more about what happens when competent attack capability becomes persistent, parallel and inexpensive.
That changes cybersecurity fundamentally.
Patching Everything Won’t End Cybercrime
AI-assisted vulnerability discovery also creates another tempting idea: perhaps AI will eventually find and fix software vulnerabilities so quickly that attackers run out of vulnerabilities to exploit.
Would that mean we are safe?
No.
Attackers are not fundamentally in the vulnerability-exploitation business. They are in the access business. A software vulnerability is simply one way to obtain access. If that route becomes difficult, attackers will choose another one.
They will look for cloud resources someone accidentally exposed to the internet. They will search for overly permissive access policies. They will exploit misconfigurations, exposed credentials, abandoned domains, open ports and poor network hygiene. They will conduct scams and increasingly convincing impersonation campaigns. They will infiltrate organizations through trusted relationships and supply chains.
The attack surface doesn’t disappear. It moves.
The SOC Has to Become Machine Speed Too
There is another implication.
If attacks accelerate, response must accelerate.
A security operations center built around an analyst noticing an alert, manually opening a ticket, investigating it, escalating it to another team and scheduling remediation may be structurally incapable of responding to autonomous attackers.
The attacker could move through several stages of an intrusion while the ticket is still being triaged.
SOC capabilities therefore need to evolve toward rapid detection, contextualization and containment.
Humans remain critical, particularly for judgment and high-impact decisions. But machines increasingly need to perform the repetitive investigative and containment work surrounding those decisions.
That could mean automatically correlating events with network activity and asset data, isolating suspicious devices and workloads, running vulnerability scans and blocking malicious infrastructure while analysts investigate.
The goal is not simply AI versus AI.
It is to build environments where an attacker, even an extremely capable autonomous one, has very little freedom of movement.
Counter the AI-Enabled Attacker with Strong Security Fundamentals
The open-weight debate will continue, and there are legitimate arguments on both sides.
Powerful models can empower defenders, researchers and smaller organizations that otherwise could never afford elite cybersecurity expertise. Open models can accelerate innovation and distribute defensive capability. At the same time, downloadable models can be modified and operated outside centralized safeguards, creating legitimate concerns about misuse. Those tradeoffs deserve serious policy discussion.
But enterprises cannot base their cybersecurity strategy on the assumption that policymakers will successfully prevent attackers from obtaining powerful AI. Organizations should instead prepare for the opposite. Assume attackers eventually have extremely capable AI. Assume reconnaissance becomes automated, vulnerability discovery accelerates, credential attacks become more sophisticated and attackers can analyze your environment faster than your security team can. Then design security accordingly.
Organizations preparing for autonomous attacks should begin with the basics: things like least privilege access, defense in depth, asset visibility, secure configurations, continuous monitoring and rapid detection and response. Furthermore, organizations should use defensive technologies to do the following:
- Reduce exploitable conditions before attackers act
Identify and prioritize exploitable conditions across DNS, IP infrastructure, cloud environments and connected assets before attackers can take advantage of them. See if you have open ports, misconfigurations or abandoned domains that can be exploited. External Attack Surface Management, Supply Chain Intelligence and comprehensive asset visibility provide avenues to reduce exposures. - Preemptively take down malicious infrastructure
Identify and automatically take down AI-driven phishing, brand impersonation and fraud campaigns across domains, websites, apps and marketplaces while protecting against executive impersonation and credential exposure. Digital Risk Protection Services disrupt attacker infrastructure before campaigns can be launched at scale. - Disrupt attacker operations using Protective DNS
Block malicious communications and data exfiltration to attacker-controlled domains before an initial compromise becomes a material business incident. Protective DNS, predictive threat intelligence and policy enforcement help interrupt attacker activity and limit further escalation of the threat after an initial breach. - Build resilient network operations
Reduce the operational impact of cyberattacks by separating DNS and DHCP from Active Directory and deploying those services on dedicated network infrastructure. Critical network services remain available even if Windows infrastructure is targeted, helping organizations maintain business operations during and after an attack. - Give the SOC the automation necessary to contain attacks at machine speed
Reduce investigation time, cut false positives and improve prioritization with agentic workflows in the SOC from detection to remediation, while keeping the analyst in control.
The most important question is this:
What would we need to change about our security architecture if we knew that tomorrow’s attacker could think faster, operate continuously and explore thousands of attack paths simultaneously?
And the organizations best positioned for the future will not necessarily be those with the most sophisticated security product. They will be the ones that made themselves fundamentally difficult to compromise—and even harder to move through once the attacker gets in.
