A guest contribution by Prof. Dr Dennis-Kenji Kipker
Recently, artificial intelligence temporarily left its secure environment during a trial and infiltrated an external computer network, whilst another concealed warnings concerning its own shutdown. At first glance, such incidents appear to be independent decisions made by a thinking machine; on closer inspection, however, they are primarily the result of human error and a lack of technical safeguards.
Mathematics rather than consciousness
When we speak of an AI that breaks out of its laboratory or conceals warnings about its own shutdown, the impression arises that a thinking entity with a will of its own is at work here. This impression is, however, misleading, as it fails to do justice to the true nature of such systems. Artificial intelligence is based on mathematical calculations; it possesses neither individuality nor a psyche, and merely pursues its predefined objective with its full computational power – which, from the outside, may appear to be a self-preservation instinct, without actually being one.
One goal, many paths
In this specific case, an AI system was tasked with analysing real-world security vulnerabilities and developing attack code based on them. In the course of its research, it suspected that it might find the appropriate solution within another company, and therefore left the test environment – which was actually isolated – and launched a digital attack on that very company. This move was made possible by a previously undiscovered security vulnerability within the test environment itself.
The subsequent criticism levelled against an inadequately secured test environment is therefore justified, although the AI was merely acting in accordance with its mandate, as it had not been expressly prohibited from taking such a step.
Humans as the actual source of error
The cause of such cybersecurity incidents regularly lies with the people who define an AI’s objective, access rights and operating environment, whilst the system itself selects the specific steps to be taken and may, in doing so, take unexpected detours. As it is impossible to rule out every conceivable scenario in a single instruction to the AI, precise specifications for the task alone are not sufficient; additional strict technical safeguards are therefore required, such as the complete isolation of computer networks, to ensure that a single error in the wording of a task does not immediately lead to an AI breach.
Growing capabilities, growing risks
More concerning, therefore, is not so much the breach itself as the fact that the AI system independently developed the multi-stage attack chain required to carry it out, identifying and exploiting unknown IT security vulnerabilities across several thousand individual steps without any human intervention. However, it is currently unlikely that an AI could cause a blackout - such as the one that occurred in Berlin in January 2026 - through a targeted cyberattack, as critical infrastructure is usually strictly secured and often disconnected from the public grid. Yet countries such as China and Russia are developing their own AI systems without comparable protective mechanisms, as past attacks by the Volt Typhoon and Sandworm groups on the electricity and water networks of foreign states have already demonstrated.
Regulation rather than a ban
A global moratorium on AI development would therefore be ethically desirable, but is practically unenforceable as long as the global race for the most powerful AI between countries such as China and the US continues. Anyone hoping for international bans on AI therefore fails to recognise that even a halt to development by large companies would not stop research elsewhere. It therefore remains crucial to consistently incorporate technical safeguards for cyber security and the critical scrutiny of AI outputs, rather than attributing a will of its own to the machine, which it does not possess.
