Security Trends 2026: 3 Learnings from Practice

Loss of Control Begins Harmlessly

A new AI tool in the team? Sounds like a manageable decision. Until it becomes access to a dozen internal systems, a prototype that goes live overnight, and shadow IT that no one fully oversees anymore. The real problem is no longer the technology itself. It’s the speed at which it’s introduced, modified, and replaced again. While IT teams are still assessing the risks of the last application, the next tool has already taken hold.

At the mid-year mark, we’re taking stock: Which security trends are we actually observing in client projects in 2026, and what should organizations prioritize now? In April 2026, Anthropic’s model Claude Mythos demonstrated what AI-powered vulnerability discovery is now capable of: Within a few weeks of internal testing, the model identified thousands of high- and critically-rated vulnerabilities in common operating systems and browsers, including a 27-year-old flaw in OpenBSD. The US Federal Reserve and Treasury Department subsequently convened bank CEOs for a crisis meeting.

From our project experience at carmasec, we draw three lessons for the second half of 2026—not as a summary of external sources, but as an assessment of what we actually see with our clients.

In Brief

• AI is increasing the pace of IT changes faster than organizations can adapt their governance. This creates new risks.
• Not zero-days, but classic basics like patch management remain the primary attack target, just with a significantly shorter response window.
• Those who embed security standards into the specifications that AI agents work with gain both speed and security.
• Concrete recommendation: Cyber resilience—from governance to technical effectiveness.

Learning 1: Immature AI Adoption Creates Vulnerabilities

The speed of IT changes is increasing faster than organizations can expand their governance to match.

It’s admirable what motivated teams in our client environment are now achieving with AI support. Prototypes emerge in hours instead of weeks, developers switch between libraries and versions in no time, new technologies are configured to run in minutes thanks to chat systems. Even the untouchability of long-established legacy systems is being challenged by a prototype created in an afternoon.

The potential is enormous!

We’re seeing a speed of change within long-established IT environments that we’ve never known before. But with it, the speed of growing risks also increases. The urge to find the next disruption and implement even more functional requirements even faster is strong. Non-functional requirements, especially security, regularly fall by the wayside. In addition to the already difficult-to-control shadow IT, there’s now an official IT that’s hardly easier to manage. AI accelerates the pace of change but fails to increase the pace of administration even remotely to the same degree.

Consequence: The same IT risks emerge as before the AI boom, just at significantly higher speed and greater scale, because administration and control don’t keep pace.

Our recommendation: Introduce governance guardrails for AI use before prototypes go live: clear approval processes, a current inventory of all AI-supported systems introduced, and regular assessments.

Learning 2: Security Basics Remain the Core Problem, Just with Less Time

Not novel zero-days are our clients’ main risk, but known vulnerabilities and misconfigurations, while the response window is noticeably shrinking.

In April 2026, initial details about Anthropic’s model Claude Mythos became known, and true to its name, truths, half-truths, and outright myths about it quickly spread. The model had examined the most secure operating systems and not only discovered vulnerabilities but exploited them independently. Such a fundamental threat to IT security, it was said, that the Federal Reserve and Treasury Department specifically invited bank CEOs to a crisis meeting.

What we observe in our own projects, however, are not attacks through spectacular zero-day vulnerabilities. What we observe are attacks that specifically search for weak configurations and unpatched, long-known vulnerabilities. It’s still the basics that attackers target, only the importance of these fundamentals has increased significantly. Patch processes today must respond within hours instead of weeks to prevent exploitation of vulnerable components.

The good news: These very fundamentals can be implemented better today than ever before through collaboration between skilled professionals and modern AI. Even without specific AI security packages, an initial assessment can often be generated quickly. Do the assigned permissions properly implement least privilege? Is the mechanism for storing credentials appropriate? Does our implementation align with common best practices? This very collaboration empowers IT teams to respond to issues more quickly.

Consequence: Patch management, system hardening, and strong detection mechanisms remain the most effective foundations for a secure organization in 2026. Consistent, rapid implementation increasingly determines risk.

Our recommendation: Streamline patch and hardening processes to hours instead of weeks and use AI specifically for daily monitoring of the basics.

Learning 3: AI + Standards = Secure Implementation

When security standards explicitly become part of the specification from which AI agents generate code, security increases despite higher speed.

The basic idea is not new. As early as 2004, Jonathan S. Ostroff and colleagues published the conference paper Agile Specification-Driven Development. The idea broke with the then-common understanding of agile software development and proposed a stronger focus on specifications. It initially remained a footnote of the conference where it was presented.

Since 2016, Sam Hatoum has shaped the term Spec-Driven Development. His manifesto is clear: specifications are the most important artifact of software development. The response remained near zero for years. Only with the emergence of AI-powered coding agents in 2025/2026 has the approach found broad acceptance. Today, the topic is meeting massive demand, whether at IBM or heise.

What has changed is the collaboration between development and AI agents. An insight from practice: AI agents find it very difficult to infer relationships and requirements that are not explicitly stated. AI agent use only becomes productive when requirements and constraints are clearly specified. Specification files fulfill exactly this need. We see this as a return to requirements and specifications documents and the waterfall model, with one crucial difference: the speed gain through AI allows this cycle to be run through multiple times a day instead of once in the project lifecycle.

Organizations that incorporate their internal security standards into these specifications experience higher security despite higher speed, in our experience. AI systems demonstrably generate more secure configurations and software when internal standards are part of the specification from which they develop.

Consequence: As a result, we see more confidence in their own security and more courage for innovation among our clients.

Our recommendation: Don’t review security requirements after the fact, but include them from the start in specifications for AI-supported development.

Conclusion: AI is changing information security in 2026 not through new attack types, but through pace

Faster introduction of new technologies, shorter response windows for known vulnerabilities, and new opportunities to embed security standards directly into development. Those who take these three learnings seriously are prioritizing now: governance guardrails for AI use, consistent patch and hardening management, and security standards firmly embedded in specifications for AI agents.

We’re happy to help leverage these opportunities without taking uncontrolled risks.