Artificial Intelligence in the Hands of Attackers: A New Generation of Threats to Critical Infrastructure
In January 2024, an employee of a financial firm in Hong Kong executed a transfer worth 25 million US dollars. He did so after a video conference with the chief financial officer and colleagues from headquarters that appeared entirely authentic. Yet on the call he was the only real person. The other participants were deepfake replicas generated by generative artificial intelligence on the basis of publicly available recordings. This case, documented by the Hong Kong police, did not announce the arrival of a new threat. It announced that the threat is already here and is operating in production mode.
The Shift from Experiment to Operations
Generative artificial intelligence has changed the economics of attack. What until recently required a team of specialists, time, and dedicated tools can today be handled by a single actor with access to publicly available models or models sold on the black market. Phishing emails in flawless Slovak, including context tailored to a specific organisation, are today cheaper than ever before. Voice clones capable of imitating a company’s executive require only a few seconds of reference audio, which can be obtained from public appearances, podcasts, or a recorded phone call.
For operators of critical infrastructure, this shift means that traditional indicators of phishing — such as grammatical errors, stylistic inconsistencies, or foreign-language context — are ceasing to work. An attack today does not need to be technically sophisticated. It is enough for it to be convincing and to reach a person who has access to the system.
Four Categories of AI-Assisted Attacks
In the threat landscape of 2025 and 2026, four categories of attacks assisted by artificial intelligence are crystallising. The first is social engineering of a new kind, including personalised phishing campaigns, deepfake voice calls, and real-time video manipulation. ENISA, in its Threat Landscape 2024, identifies deepfake attacks as one of the fastest-growing categories, with a year-on-year increase of several hundred percent.
The second category is automated reconnaissance and exploit development. Large language models are able to process publicly available documentation, configuration files leaked from public repositories, and source code, and to identify potential vulnerabilities from them. Genuinely new exploits still require expertise, but the reconnaissance phase, which forms a significant part of the attack lifecycle, has been substantially accelerated.
The third category is adaptive malware capable of changing its behaviour according to the environment in which it lands. Generative models make it possible to create polymorphic malware variants that bypass signature-based detection systems. For the time being, these are primarily experimental demonstrations, but the line between research and operational deployment is growing thinner.
The fourth category is content manipulation and disinformation, which has a direct impact on critical infrastructure through public trust. False reports of an accident at a nuclear power plant, contamination of drinking water, or an outage of payment systems can trigger mass panic even before the operator manages to publish accurate information.
AI as a Component of Critical Infrastructure
The second dimension of the problem is that artificial intelligence is itself becoming a component of critical infrastructure. AI systems are being deployed in energy-sector dispatch centres for consumption forecasting, in healthcare for clinical decision support, in the financial sector for fraud detection, and in transport for traffic flow management. Each of these systems becomes a new attack surface. Poisoning of training data, real-time manipulation of input data, and model-inversion attacks are categories of threats that traditional cybersecurity tools cover only to a limited extent.
The European Union’s Regulation on Artificial Intelligence (the AI Act), which entered into force in August 2024 and is being progressively applied through 2027, classifies AI systems in critical infrastructure as high-risk, with obligations covering risk management, documentation of training data, transparency, and human oversight. Operators of critical infrastructure thus find themselves at the intersection of three regulatory frameworks: Act No. 367/2024 Coll. on critical infrastructure, Act No. 366/2024 Coll. on cybersecurity, and the AI Act.
Three Priority Areas for Operators
Three priority areas follow for operators of critical infrastructure. First, the reassessment of identity-verification processes in communication via voice and video channels. What was sufficient in 2020 is not enough today. Telephone instructions for financial transfers, changes to access rights, or operational interventions should be moved to verification procedures resilient to deepfake attacks, including call-backs to verified numbers and multi-channel authentication.
Second, lifecycle management of the AI systems deployed in operations. An inventory of the models in use, documentation of training data, monitoring of deviations in outputs, and procedures for switching to manual mode upon detection of an anomaly are today part of cyber hygiene, not an add-on.
Third, training of personnel with an emphasis on the understanding that an attack may come in a form an employee has not previously encountered. Regular exercises that simulate deepfake calls from company management or AI-assisted phishing campaigns are more effective than traditional awareness training based on text-based examples.
“Artificial intelligence does not bring an entirely new category of threats. It brings a new pace and a new accessibility for those we have already known. An attack which only three years ago required a team and weeks of preparation can today be prepared by a single actor in an hour. This is a shift to which it is not enough to respond with better antivirus systems. It calls for a reassessment of processes, the human factor, and verification regimes in every critical entity,” says Ing. Tibor Straka, Chairman of the Critical Infrastructure Association of the Slovak Republic.
The security of critical infrastructure in the era of artificial intelligence will not be about a race in technology. It will be about whether operational processes are able to adapt faster than the actors on the other side.










