Once handed an attack target, the artificial intelligence began scanning the environment. It probed externally exposed ports and services and searched for entry points. Based on the information it gathered, it drew up an infiltration plan and selected the tools and commands it needed. Moments later, "FAIL" appeared on the screen, but the attack did not stop. The AI analyzed the cause of the failure, attempted infiltration again by another method and switched routes when blocked. Without a human issuing the next command, it repeated a cycle of scanning, planning, execution, analysis and retrying, finding its own way through.
This was a demonstration of an "AI penetration test" released by AhnLab on the 16th of last month. After several attempts, the AI found a vulnerability in a file upload process and confirmed the possibility of remote code execution, or RCE, which allows arbitrary commands to be issued to a server from outside. What sets it decisively apart from conventional automated attacks is what happens after failure. Rather than simply repeating pre-programmed attacks, it looked at the results and decided its next move.
At the time, AhnLab presented this as a new form of attack that would emerge in the AI era. Only weeks later, those concerns materialized in South Korea's financial sector.
Concerns are mounting that AI-driven cyberattacks could spread beyond the financial sector to critical national infrastructure, according to the security industry on the 5th. Power, telecommunications and transportation infrastructure has stronger security systems than the financial sector, but the threshold for mounting an attack is falling as AI takes over everything from scanning and vulnerability discovery to selecting attack methods and analyzing failures before retrying. When AI rapidly repeats a process that once required attackers to analyze system architecture piece by piece and hunt for weaknesses, it can uncover external touchpoints and back doors that were previously hard to find. In particular, as the time, cost and expertise required for an attack decline, critical national infrastructure that was once out of reach because of high security barriers could come within the scope of attackers' searches, analysts said.
Lee Seung-kyung, head of AhnLab's AI development division, said, "If AI agents lower the cost of attacks and the technical barriers to entry, attempts targeting national infrastructure are likely to increase as well." The executive added, "Since AI can quickly find weak points and attempt attacks on a large scale, response systems that preemptively identify and block vulnerabilities will become even more important."
Attempts to use AI against critical national infrastructure have already surfaced overseas. According to Dragos, an industrial control systems security firm, an attacker with no expertise in operational technology or ICS attacked a water utility in Mexico early this year using Anthropic's Claude and OpenAI's GPT models.
Claude was used for technical execution, including infiltration planning and the development and deployment of attack tools, while the GPT models handled processing of collected data and structuring of analysis results. Claude, in particular, carried out a large-scale automated attack that indiscriminately tried passwords against the relevant interface, much like the recent hacking case in the financial sector. The attempt ultimately failed, but it showed that even attackers lacking expertise can use AI to find attack paths into critical infrastructure.
In August this year, the U.S. National Security Agency, together with the Cybersecurity and Infrastructure Security Agency and the Federal Bureau of Investigation, warned of cyberattacks targeting Siemens' S7 series programmable logic controllers. Attackers are using AI-generated exploit scripts, the agencies said. PLCs are core devices that automatically control equipment at factories and other industrial facilities.
The NSA warned that an attack on a poorly secured PLC could not only halt key industrial processes but also lead to "real-world effects" such as safety incidents, equipment damage and shutdowns. The potential targets extend well beyond core manufacturing to industries directly tied to national operations and daily life, including energy production and distribution, water and wastewater treatment, chemicals, and food and agriculture.
While AI attacks on the financial sector damage information and financial systems, cyberattacks on critical national infrastructure can translate directly into physical harm. Failures in power generation and transmission or in water and sewage systems directly affect the operation of factories, data centers and hospitals as well as public safety, and a halt to railways, aviation or ports leads to logistics disruptions. As policing, firefighting, disaster response, military command and control, and logistics systems become increasingly digitized, public safety and defense cannot be exceptions either, observers said. This year alone, a cogeneration plant in Poland, Brussels Airport in Belgium and a water system in Minnesota in the United States have all been revealed to have suffered hacking damage.
As AI increases the speed and scope of attacks, cybersecurity is turning into a race over who finds a vulnerability first, the attacker or the defender, analysts said. Experts say critical national infrastructure also needs a preemptive security framework in which defenders find and block weaknesses before attackers locate them with AI.
Lee Sang-geun, director of Korea University's Institute of Artificial Intelligence Security, said, "Critical national infrastructure such as power and defense has a higher level of security than the financial sector, but there may be external touchpoints or vulnerabilities that have not yet been identified." The director added, "When cyberattacks on critical infrastructure became a reality in April, the U.S. government also carried out a full audit of national infrastructure." The director stressed, "We, too, need to conduct a full audit of the assets and touchpoints of all systems and services and use AI-based penetration testing to preemptively find and block hidden back doors."
