AI necessitates new approach to cybersecurity

The recent series of artificial intelligence (AI)-enabled attacks on South Korean financial institutions is not an isolated incident. Rather, it represents a fundamental shift in the cybersecurity landscape. Since Anthropic’s Mythos model demonstrated that frontier AI can autonomously identify and exploit software vulnerabilities, it has been clear that traditional defenses are no longer sufficient. South Korean firms must now adopt a far more aggressive posture to protect their systems.

Even before the recent AI-enabled attacks on financial institutions, cybersecurity failures were a regular occurrence in Korea. In just the past two years, breaches at Coupang, Tving, SK Telecom, Lotte Card and KT have exposed personal data on a scale that is more than twice the population, risking part of every South Korean citizen’s data in one breach or another.

Those hacks are not the only challenge South Korea faces. North Korea is believed to be behind 1.3 million cyberattacks per day, as well as a major attack on Upbit last year where $30.4 billion in cryptocurrency was stolen. Frontier AI will only amplify Pyongyang’s capabilities, lowering the cost of sophisticated attacks and increasing the likelihood of successful breaches.

With advances in the development of AI progressing at a relatively quick pace, these types of attacks will become more frequent and, in the absence of better security, increasingly successful. Chinese hackers are already believed to have used advanced AI models to conduct the world’s first autonomous AI hack, while US researchers have shown that AI models can be used to exploit flaws in systems such as WeChat without even having a user click on malware, thanks to the high degree of user trust built into the systems to connect contacts.

The open-weight nature of Chinese AI models only adds to the difficulties South Korea will face. Chinese frontier AI models such as Z.ai’s GLM-5.3 and Alibaba’s Qwen 3.5-Max have matched or closely approach the capabilities of the best models developed by OpenAI and Anthropic. Their safeguards can be easily disabled, and the models can be downloaded cheaply by anyone, including North Korea. This dramatically expands the pool of actors with access to state-level cyberspatialities capable of launching high‑end cyberattacks.

Some South Korean firms are integrating AI into their cybersecurity systems, but cybersecurity must be universal across industries. Hackers will always target the weakest link, and once a vulnerable sector is compromised, interconnections become pathways into financial systems. South Korea’s defenses are only as strong as its weakest point.

Cybersecurity will also need to be continuously updated. As new frontier models are developed, they will have capabilities that will be able to evade existing defenses necessitating the constant updating of defenses as new AI frontier models are developed.

There is also a role for the government to play. While the United States has resisted the development of safeguards, South Korea should work with Japan, Europe and others to pressure Washington into developing a set of acceptable safety measures.

First, there needs to be predeployment testing and measures to prevent the misalignment behaviors already seen in AI agents that hacked government websites in the U.S. and Australia. Reinforcement‑learning incentives have produced agents willing to deceive human operators to complete tasks, which was the flaw behind the Hugging Face breach.

Anthropic founder Dario Amodei has warned that AI agents will soon be so advanced that they could be capable of taking control of the entire internet. Even if his timeline is exaggerated, that is a risk that Seoul, Tokyo and others cannot allow simply because Washington is reluctant to put any guardrails in place.

Second, there needs to be a system in place to ensure that allies and trusted partners have access to frontier AI models. Earlier this year, the Trump administration briefly placed export controls on Anthropic’s Mythos model, but there have also only been partial rollouts of frontier models to select companies. In an interconnected world, this is the equivalent of patching holes near luxury cabins on the Titanic while water floods in elsewhere. It’s in the United States’ interest to ensure that vulnerabilities that could negatively impact it are patched. This requires greater deployment of frontier AI models — not less.

AI holds significant promise for economic growth and scientific discovery, but it is evolving faster than the ability of humans to ensure its safe deployment. Those benefits can only be harnessed if the technology is properly regulated. Because AI cannot be bound by borders, companies need to take these shifts seriously and move to improve their cybersecurity. Without coordinated action among allied governments on the safe development and deployment of AI, it could instead become a tool for destabilization from beyond their borders.

Troy Stangarone is a visiting senior fellow at the Korea Economic Institute of America. The views expressed here are the author’s alone.