Attackers work at machine speed. What does that mean for your security approach?

For CISOs, IT managers and security officers who have to explain digital resilience to management, regulators and customers.

Attackers become faster, smarter and more convincing. According to the Microsoft Digital Defense Report 2025 , AI-created phishing emails are clicked on much more frequently than traditional campaigns. Where traditional phishing campaigns average 12 percent clicks, AI-generated phishing has reached 54 percent. There is often a clear revenue model behind this. More than half of the attacks for which the motive is known involve extortion or ransomware. And in many cases, it's all about data: in 80 percent of the incidents investigated, attackers were out to steal data.

AI thus makes cybercrime not only more convincing, but also more scalable. Attackers can reconnoiter faster, deceive more effectively, and execute attacks at a faster pace.

For organizations, this means more than a new generation of phishing emails. Attackers are increasingly moving at machine speed. Those who still rely entirely on human speed in detection and response are structurally lagging behind. At the same time, that same technology also offers opportunities for defense. AI helps to analyze, make connections and intervene faster when threats develop.

This makes it increasingly important to make AI a safe, transparent and controlled part of your security operation.

What will change now that attackers are using AI?

There are three developments that stand out.

• The first is automated social engineering. With AI tools, attackers can map out how an organization works at lightning speed. Which departments are there? Who works with whom? Who emails about what? They use that information to create phishing campaigns that can hardly be distinguished from the real thing. Including the right tone, names and colleagues in cc.
   
• The second development is speed. A ransomware attack that used to unfold in days can now cause damage in minutes. Detection and response must be able to keep up with that pace. Otherwise, help will simply come too late.
   
• The third development is the automatic finding and use of vulnerabilities. Powerful AI models show how fast this development is happening. For example, in the white paper AI and arms race , we mentioned Mythos-class models: systems that can identify and exploit software vulnerabilities autonomously, without human intervention.

That does not mean that every attacker will have such capacity tomorrow. It does show which way it is moving. As AI gets better at recognizing vulnerabilities and combining steps in an attack chain, poorly patched or outdated environments become interesting more quickly. What now seems exceptional or controlled can come closer to practice in a short time.

Defending AI-first: what does that mean in practice?

AI-first security means that AI is not an extra layer on top of your service, but becomes part of the foundation. From detection to analysis and response. Within our MxDR service, Managed eXtended Detection & Response, this takes shape in three ways.

1. First of all, we use AI deterministically. The analysis process remains leading and predictable. AI is used in a targeted way for steps that cannot be captured well in fixed if-then rules. Think of recognizing anomalous patterns in a series of data points. In this way, you combine the reliability of a proven process with the analytical power of AI, without the outcome becoming a black box.
    
2. In addition, security must be able to respond to machine speed. Technology such as automatic attack disruption within the Microsoft security platform can automatically intervene when, for example, a ransomware attack begins to unfold. In combination with automation in the Cyber Defense Center, we reduce the response time to the pace at which the attack occurs.
    
3. And perhaps most importantly: people remain in control. AI helps analysts to make connections faster and make complex incidents understandable. Even for people who do not work in security jargon on a daily basis. Decisions with impact stay with people. This creates a direct connection between what has been agreed administratively and what happens operationally when a threat occurs.

The attack surface is growing: from mailbox to API and AI agent

AI isn't just changing the pace of attacks. It also changes what needs to be protected.

The mailbox remains the front door. Of course, automated filtering helps, but user reports also deserve serious follow-up. An e-mail that is reported as phishing by an employee should not get stuck somewhere in a queue at system administration. That report should end up in a security operation that analyses, interprets and intervenes before the same e-mail causes damage to the next colleague.
 

APIs are becoming increasingly important. Almost every application will be accessed via an API in the coming years, especially now that AI agents increasingly need to be able to work with systems. That traffic must be demonstrably safe and legitimate. By monitoring API traffic for anomalies, attacks, and suspicious signals, the API gateway becomes a guarded access rather than a blind spot.

AI agents also require their own security. Let's say you use a customer service agent on your website. Then you want that officer to continue to respond properly, not leak data and not be fooled with malicious prompts. This requires a control layer that tests both the input to the agent and the output of the agent.

Data deserves protection at the source. Employees use AI tools, with or without permission. With data classification, you make visible which information is shared with which AI services. You can then enforce policy on that. For example: sensitive data may go to your own Microsoft Copilot environment, but not to external AI tools.

Finally, it pays to look at your organization as an attacker does: from the outside. With eXtended Attack Surface Management, you can map out which IP addresses have been unlocked, which ports are open, which software is running behind them and whether that software has been patched. The inside in order, the outside in the picture.

From tooling to provability

Boards, supervisors and auditors are increasingly asking a different question. Not: do we have a SOC? Or: do we have tooling? But: can we demonstrate what happened, how we reacted and why that was the right response? Laws and regulations such as NIS2 and the Cybersecurity Act reinforce this movement.

Provability requires transparency in the security operation itself. Think of reports in which detections are linked to the MITRE ATT&CK framework, so that it becomes clear where in the attack chain threats are recognized. Think of an audit trail per incident: which actions were taken, by whom and with what result. And think of interpretation in understandable language, so that accountability does not depend on that one colleague who can translate all the technical details.

Start with factual insight

If you want to know where the organization stands, you don't start with tooling. You start with factual insight.

For example, ask yourself these questions:

• Which AI tools do our employees use and what data do they share with them?
• Which APIs and AI agents are externally accessible and who monitors that traffic?
• How quickly do we detect and stop an attack that unfolds in minutes?
• Can we demonstrate to the management and supervisor how an incident was handled?

The answers show where the base is, but also where the attack surface is growing faster than the grip on it. That insight is not an end point. It is a starting point for making better choices. From sharpening data classification to having your attack surface scanned externally.

Would you like to talk further about what AI-first security means for your organization? Our security specialists are happy to think along with you. We start with insight into where you are now.