AI Attackers Are Getting Faster. Are Your Cyber Defences Still Waiting for a Human?

Is your cybersecurity ready for the speed of AI attackers? Discover how to enhance your defenses against automated threats.

Article

Imagine your company's security monitoring system detects something unusual. An employee account has signed in from an unexpected location, and shortly afterwards, it begins accessing information that the employee wouldn't normally use. The monitoring software generates an alert, which is passed to the security team for investigation.

Someone needs to determine whether the activity is legitimate, establish what might have happened and decide whether the account should be restricted.

In most businesses, that process involves people. An analyst reviews the evidence, perhaps checks the employee's recent activity, contacts another team or asks for approval before taking action. There's a good reason for that. Disabling the wrong account or interrupting an important business service can create its own problems, particularly when the original alert turns out to be harmless.

But I've been thinking about what happens when the suspicious activity is progressing faster than the organisation can investigate it. The security team might be following the correct process, making sensible decisions and taking appropriate care, yet still find itself responding too slowly to prevent a relatively small incident from becoming something more serious.

That question has been on my mind since reading Booz Allen Hamilton's September 2026 report, *The Offensive Frontier: AI as the Attacker*. The research examined how advanced AI models performed when attempting cybersecurity tasks in a controlled enterprise-style environment.

Some models demonstrated the ability to progress through multiple stages of an intrusion without needing a person to guide every action, suggesting that increasingly capable systems could change the speed and effort involved in certain types of attack.

The researchers weren't claiming that every AI model can compromise any organisation, or that conventional security controls have suddenly become ineffective. Real networks differ considerably, and the experiments took place under defined conditions. Nevertheless, the results raise an important operational question for businesses that are already struggling to keep up with security alerts, complex systems and limited internal resources.

If AI allows attackers to automate more of the work involved in finding and exploiting weaknesses, can organisations continue relying on response processes that require people to approve every step?

I don't think the answer is simply to automate everything in return. In fact, doing that without understanding the consequences could introduce an entirely different set of risks. The more useful question is which defensive activities can be performed reliably and safely through automation, and where human judgement still provides an essential safeguard.

## When cyber threats move faster than the response process

Cyberattacks have involved automation for many years. Software can already scan networks, identify exposed services, attempt to misuse compromised credentials and carry out repetitive activities at considerable speed. So I wouldn't suggest that artificial intelligence has suddenly introduced automation into a previously manual activity.

What's changing is the range of tasks that increasingly capable AI systems may be able to coordinate and the way they can respond when an initial approach doesn't work.

Traditional automated security tools generally operate according to defined rules or sequences of actions. An AI-supported system may be able to interpret information from its environment, adjust its approach, use different tools and continue working towards a broader objective. That doesn't mean it operates with the judgement of an experienced cybersecurity professional, or that it will reliably complete every task.

But it could reduce the amount of direct human involvement required for certain activities that previously demanded specialist attention.

Booz Allen's Cyber Weapon Index was designed to examine some of those capabilities. In its September research, the company evaluated 18 advanced AI models across tasks involving vulnerability research and progress through a controlled network intrusion. The researchers used observed activity and system evidence to assess what the models actually accomplished, rather than relying solely on their descriptions of success.

The results showed substantial differences between models. Some made limited progress, while a leading system completed the defined intrusion sequence under the test conditions. The research also found that supporting software, including tools, memory and mechanisms for coordinating activity, could significantly increase what a model was able to achieve.