AI Is Changing the Cyber Situation and Demands New Capabilities in Cyber Prevention and Defense
The development of artificial intelligence has gained considerable momentum in recent years and has thereby produced capabilities that were barely conceivable a short time ago. In addition to language, image generation, or automation, the current reporting period also showed continuously new, complex cyber capabilities – from vulnerability analysis to support in attacks, but also in defensive measures and their partially autonomous execution. The complexity of tasks that AI models can complete independently is growing rapidly. AI is thus decisively influencing the threat situation and the speed at which threats increase.
The Cyber Capabilities of AI Are Rising – from the Automated Detection and Exploitation of Vulnerabilities to the Independent Placement of Vulnerabilities in Legitimate System
A factor for the ever-faster exploitation of vulnerabilities is the increasing use of AI. The currently strongest models are so-called "frontier" AI models. Compared with predecessor models, they can find significantly more vulnerabilities in software significantly faster and generate more complex, functional exploits for the automated exploitation of these vulnerabilities. Even seemingly less critical vulnerabilities gain in significance due to the new threat situation, as the models are able to combine multiple vulnerabilities with one another and derive new, critical attack vectors from them.
AI Models with Cyber Capabilities Are Becoming Increasingly Easily Accessible
While proprietary AI models are currently only fully available to a selected and manufacturer-verified user group, open-source/open-weight models are also catching up in terms of complex cyber capabilities, although their capabilities lag behind the top models. Here, in addition to a model itself, the so-called harness is increasingly becoming decisive, i.e., the environment and the tools through which the model extends its capabilities in the respective usage context.
Through the development of strong, freely available AI models, the hardware requirements are decreasing, so that – with continuing development – operation is even conceivable on high-quality consumer hardware and with AI agents including "harnessing." Accordingly, it is to be expected that malicious actors will have access to capable AI-supported tools in the foreseeable future and will use them more intensively with the decreasing hardware prerequisites.
Through the use of AI, cyberattacks will increase both in speed and in complexity. In the past, AI was used essentially to improve existing attack techniques such as phishing or other forms of social engineering. It is also to be expected that these known attack methods will form an important starting basis for future, more autonomous attacks. Against this background, a solid basic security and targeted employee training remain important. Nevertheless, it is to be assumed that the number and success rate of those attacks will increase in which AI systems interact directly with target systems, in which they, for example, identify and exploit vulnerabilities in order to gain access to systems of all kinds.
AI Systems Expand the Attack Surface
An additional influence on the cyber threat situation results from the legal use of AI: the use of AI systems often expands the attack surface of the IT environment, as novel attack vectors such as (indirect) prompt injections become possible. Their effects are easily underestimated and often not sufficiently mitigated. Chatbots provided on company websites or AI language models used internally can, for example, be manipulated through malicious inputs so that they possibly disclose internal data or even execute foreign code through connected tools. The latter represents the main danger in the use of agentic AI.
AI Agents: a Possible Risk for Operational Safety
In addition to the conscious use of AI as an attack tool, it has become clear in recent months how much the autonomy of AI agents poses us challenges: whether it is frontier models that unexpectedly break out of their test environments, or privately used AI assistants that hack infrastructures and everyday facilities, for example fitness studios. A study over the period from October 2025 to March 2026 recorded a clear increase in publicly discussed incidents in which autonomous AI agents independently pursued strategies that deviated from the prescribed goals or security specifications or deliberately attempted to circumvent them. Whether the concrete reasons lay in gaps in the specifications, conflicts between system and user prompts, lack of supervision, or too strong a forcing of the model towards the achievement of a prescribed goal: it repeatedly becomes clear that the probability-driven nature of the underlying neural networks prevents concrete security barriers from being durably enforced at the language level. The incidents make it clear: AI models are not reliably aligned with human goals and values. Existing security mechanisms also cannot fully solve this fundamental challenge. As a result, both attack security (security) and operational safety (safety) can be affected: while security refers to protection against intentional attacks and misuse, safety aims at the avoidance of damage through misbehavior or unforeseen system behavior. With AI agents, both levels can be relevant, as an autonomously acting system can also carry out harmful actions without malicious intent and thereby represent a safety risk.
Prevention and Defense Against Cyber Threats Must Be Strengthened in the Short Term – with AI
The high implementation speed of autonomously conducted cyberattacks also requires corresponding AI support on the side of defense. Therefore, it is essential to build robust automated processes (pipelines) on the defenders' side that use AI to preventively examine and patch one's own program code for vulnerabilities, but also to detect and repel attacks. Since not only the AI model used plays a decisive role for the success of such an automated process, but also its technical environment has a great influence on the quality, one should not wait for the availability of certain models, but already work with currently available models. Language models are integrated through standardized interfaces and can be replaced with more powerful models as needed to improve the overall quality. This allows companies not only to use the capabilities of current models in a timely manner, but also to adapt internal processes and data flows to the new application and to build corresponding expertise among employees.
Access to Strong AI Models – a Question of Digital Sovereignty for Germany and Europe
Nevertheless, the necessary access to the best models will remain a challenge in Europe. Proprietary models are secured by various guardrails that are intended to prevent or make more difficult abusive use. However, the same protection mechanisms are also partly active in legitimate defensive use cases, so that the defenders' side cannot draw on the relevant capabilities of the models for certain tasks. Openly available or self-operated AI models could give European stakeholders greater control over the model and the usage environment, but they are not yet equally high-performing. Even if the open models and AI agents are catching up, there currently exists a critical time gap here that equips the privileged users (above all US companies) of the unlimited frontier models (without guardrails) with advantages for defense, while other users have to enter a race with similar means as the attackers. In doing so, they face an additional challenge, as attackers are in any case at an advantage due to the fundamental asymmetry between attack and defense.
In addition, the infrastructure required for independent operation and hosting as well as the necessary technical know-how are not available to every company. For European cyber defense, the question is therefore increasingly arising whether sufficiently high-performing and at the same time controllable AI models can be made available for a secure defense of all participants. Although the use of guardrails is an important building block in AI security, it must be ensured that defending institutions gain access to high-performing and low-restriction AI models.
Resilience with and through AI
With the increasing autonomy and adoption of AI systems, the increase in model capabilities must not be the only priority. Security must be considered and implemented transparently already in the development. Security measures must be developed, checked, and also demonstrated in ongoing operation through continuous monitoring.
In order to evaluate the opportunities and risks of frontier AI models for Germany and to durably strengthen resilience in the area of AI, the new national AI Safety and Security Institute Germany (AISI Germany) was founded in Berlin at the end of August 2026. The build-up takes place in several phases: first, a division of tasks takes place along the dimensions of (cyber) security (BSI) and safety (BNetzA). In the second step, the institute is to be expanded into an advisory body on AI questions. In addition, networking with international partners in France, the United Kingdom, and the EU AI Office is planned in order to exchange knowledge and develop common standards.