[Gaps in Financial Security] 7 "AI Attacks Must Be Countered with AI"... How Are Financial Institutions Abroad Responding?

IMF and BIS: Detection, blocking, and recovery must speed up to keep pace with AI attacks
The UK tests multistage attacks; Singapore conducts AI red teaming
South Korea introduces AI red teaming for financial-sector penetration testing for the first time

As cyberattacks using artificial intelligence (AI) become more sophisticated and automated, financial institutions are increasingly adopting the approach of “using AI to defend against AI attacks.” Rather than viewing AI solely as a new security threat, they are deploying it across their entire defense systems—from vulnerability detection and simulated attacks to incident response and recovery—to keep pace with the speed of attacks. There are calls for South Korea’s financial sector to accelerate the integration of AI throughout its security systems.

According to the report “Artificial Intelligence and Cybersecurity in the Financial Sector,” published by the International Monetary Fund (IMF) in June, major advanced economies are moving toward using AI in cybersecurity. They are applying AI to vulnerability detection and analysis, as well as response and recovery, to narrow the speed gap between attacks and defense. The IMF said that as AI makes cyberattacks faster and larger in scale, it is creating an environment of “machine-speed” attacks and defense that is difficult for humans to keep up with. Rather than focusing only on preventing intrusions in advance, the entire process needs to be enhanced—from rapidly detecting attacks and containing the spread of damage to restoring systems.

Countries are already moving quickly. The UK AI Safety Institute (AISI) is using cyber ranges to test whether the latest frontier AI systems can autonomously identify vulnerabilities and carry out multistage intrusions like real-world attackers. The Bank of England (BOE) is also working to incorporate AI-related scenarios into cybersecurity and operational tests for the financial sector.

The U.S. Federal Reserve (Fed) also believes that generative and agentic AI could introduce new risks to the security of systems operated by financial institutions, financial infrastructure providers, and third-party service providers, and is reviewing its supervisory framework. The Fed is considering whether it is appropriate to apply existing risk management frameworks unchanged to banks’ use of generative and agentic AI, while also revising its standards for managing third-party service risks.

Japan has launched a public-private partnership that brings together financial institutions, AI and IT companies, the government, and the central bank to address cyber threats posed by frontier AI. Participants include financial institutions such as Seven Bank, Mizuho Bank, and MUFG Bank; IT companies such as Amazon Web Services (AWS), Anthropic, OpenAI, Google, Microsoft (MS), IBM, NEC, and Fujitsu; as well as the Japan AI Safety Institute, the Ministry of Finance, the Bank of Japan (BOJ), and the Financial Services Agency (FSA).

Singapore’s example is noteworthy because it shows how AI is being used for defense in an Asian financial sector like South Korea’s. Since July 1, the Monetary Authority of Singapore (MAS) has required major financial institutions to conduct “AI-assisted red teaming” on critical internet-facing systems. The approach uses advanced AI models to identify attack paths that cyberattackers could exploit to disrupt critical services or access sensitive customer information. MAS plans to share vulnerabilities identified and response examples from the exercises with the financial sector to raise the level of defense across the industry.

While Singapore is requiring major financial institutions to conduct AI red teaming, South Korea is also stepping up efforts to apply AI to financial sector security. This year, the Financial Supervisory Service and the Financial Security Institute introduced AI red teaming for customer-facing AI services for the first time as part of the financial sector’s blind penetration-testing exercise. The Financial Security Institute has established a dedicated AI red teaming team and is also strengthening its assessments of AI agents.

The AI initiatives of Japan’s SBI Holdings and Visa are also worth watching. SBI Holdings is working with Anthropic on joint testing to apply “Claude Security,” the latest AI security technology, to the financial sector. The company plans to check for security vulnerabilities in generative AI services under development at SBI Securities and other affiliates, and to establish an AI security framework suited to the financial sector.

Visa, one of the world’s largest payment network operators, uses global transaction data and hundreds of AI models to detect fraudulent transactions. Visa said that in 2024, its security systems, which use hundreds of AI models and other technologies, blocked fraudulent activity worth $40 billion (about KRW 55 trillion).

South Korea is at the stage of expanding the institutional foundation for using AI as a defense tool. Since June, financial authorities have temporarily eased network separation regulations for some financial institutions when AI is used for security purposes. In the first round, 10 financial institutions began security testing using high-performance AI, and the authorities confirmed that AI could analyze vast amounts of source code in a short time and identify vulnerabilities across a wide range of systems.

However, the financial sector as a whole has not yet reached the point of actively using AI as a defense tool. Following widespread hacking incidents targeting the financial sector, including a recent attack on Shinhan Bank, the selection of financial institutions for the second round of eased network separation regulations, initially scheduled for October 7, was postponed. This was because financial authorities’ personnel were focused on responding to the incidents and conducting urgent security checks.

A Financial Services Commission official said, “The government’s security policy and stance of responding to AI threats with AI have not changed. However, as we focused on responding to the hacking incidents, we were unable to complete our review of the candidates for selection. Once the situation is under control, we plan to speed up the selection process.”

Kim Seung-joo, a professor at Korea University’s Graduate School of Information Security, said, “Frontier AI is not necessarily required to defend against hacking. It is possible to mount an effective defense by combining multiple commercially available AI tools, and related tools and products are already on the market.”

However, he stressed, “At present, network separation is not complete, making it difficult to use AI across all areas of cybersecurity. Since there are also constraints on conducting red teaming exercises, complete network separation must come first if AI is to be used comprehensively for security.”

This content was produced with the assistance of AI translation services.