Ten Korean Financial Firms Failed Security Checks Before Connecting Networks Under AI Exemptions

Inspections found safeguards existed mainly on paper

Ten South Korean financial companies granted temporary permission to connect internal business networks with external internet services had failed to put many required security safeguards into operation, inspections conducted from mid-July to early September found. The deficiencies included security management consoles accessible from the internet and a cloud service without security certification.

The Financial Security Institute reviewed 12 alternative security controls after the Financial Services Commission granted the firms a one year exception on June 19. The permission was intended to let them use external artificial intelligence and security software to identify vulnerabilities in internal systems.

The institute said all deficiencies were corrected before the internal and external networks were connected. That distinction matters: the findings show inadequate preparation for the exemptions, but do not establish that hackers entered internal systems through connections authorized by the programme.

The inspection materials were obtained by Rep. Han Chang-min of the Social Democratic Party, a member of the National Assembly's National Policy Committee. They showed that many safeguards remained at the planning stage and that firms had not verified whether their controls worked in their own operating environments.

The findings emerged as regulators delayed a second round of exemptions that would have selected up to 15 institutions. Authorities attributed that postponement to the need to concentrate administrative resources on a separate wave of breaches and data leaks, rather than announcing an end to the policy of relaxing network separation rules.

Which companies were inspected?

The first group consisted of four banks, three securities companies, two insurers and one card company. The banks were Shinhan Bank, Hana Bank, Woori Bank and KakaoBank. The securities firms were KB Securities, NH Investment & Securities and Mirae Asset Securities. Samsung Fire & Marine Insurance, Hanwha Life and Hyundai Card completed the group.

All ten were subject to checks on the alternative controls required for their exemptions. The institute did not disclose which company had which deficiency, or publish a detailed allocation of the failed control items, citing the risk that such information could help attackers identify targets.

Consequently, the finding that some administrator consoles were accessible from the external internet cannot be attributed to any named participant. The same applies to the uncertified cloud service. Listing the inspected firms identifies the scope of the review, not responsibility for each specific weakness.

The institute reported its findings to the Financial Services Commission and said it confirmed remediation before network connections proceeded. No customer loss, successful intrusion or data leak caused by these particular deficiencies was identified in the inspection account.

What the temporary permission required

Financial companies generally must separate internal business systems and information processing systems in their data centres from external communications networks, including the internet. The June 19 measure temporarily relaxed that boundary for a specific purpose: using external AI and security software as a service to check internal systems for vulnerabilities.

Software as a service, or SaaS, is software accessed through a provider's online infrastructure instead of being installed and operated entirely within a company's own systems. Using it to inspect internal technology can require a controlled connection across a boundary that network separation rules would otherwise restrict.

The exception was therefore conditional. Firms had to establish and implement 12 substitute controls, including procedures for responding to security incidents and disruptions involving AI or SaaS services. They also had to prevent unnecessary sharing or processing of data within those services.

These conditions addressed more than the ability to connect safely. They also covered what happens when an external provider stops working and how information is handled once an outside service becomes part of a firm's security operations. A written response plan could satisfy a planning requirement without demonstrating that staff, systems and access restrictions were ready to execute it.

Why exposed administrator consoles matter

The inspections identified a common failure to verify controls in the environments where they would actually operate. Some firms also left security administrator systems, known as management consoles, accessible from the internet. These interfaces are used to administer security tools, making restrictions on access especially consequential.

Internet accessibility does not by itself prove that an attacker bypassed authentication or gained control. It does, however, create an external route to a sensitive administrative interface. The inspection disclosures do not specify the authentication arrangements, duration of exposure or number of affected institutions.

Yeom Heung-yeol, a professor in the Department of Information Security at Soonchunhyang University, warned that the safeguards might not have operated fully before corrective work was completed. He singled out access to security management systems as a particularly serious deficiency:

Failing to restrict external access to the security administrator system is a very serious matter.

The institute's account places remediation before the authorized connections. That limits what can be concluded about actual exposure of internal networks, while leaving a substantial compliance concern: controls intended to justify an exception had not been fully implemented when inspectors checked them.

A second round paused amid separate breaches

The Financial Services Commission had planned to select the next group of institutions on October 7. The postponement was reported on October 6, with authorities citing the need to focus resources on recent hacking and data leakage incidents and to monitor developments before expanding the programme.

Incidents had been reported since September 27 at institutions including KB Kookmin Bank, Shinhan Bank, Hana Bank, BNK Busan Bank, Yegaram Savings Bank, Welcome Savings Bank and Hyundai Capital. The reports do not establish that those incidents resulted from the temporary network separation exemptions.

Shinhan Bank and Hana Bank appear in both the inspection group and the list of institutions with reported incidents. That overlap is not evidence of a causal connection. The security review and the wider breach response must be distinguished unless further findings connect them.

The main dates describe a sequence of permission, inspection and delayed expansion:

• June 19: The commission granted ten firms temporary exceptions for external AI and security SaaS use.
• Mid-July to early September: The Financial Security Institute inspected the firms and required corrections before network connections proceeded.
• September 14: Applications for the planned second round were scheduled to close.
• From September 27: A series of financial sector breaches and data leaks was reported.
• October 6: The postponement of the October 7 selection was reported.

Authorities said the delay did not change their position on easing rules for firms with adequate response capabilities. No replacement selection date was given.

The expansion would reach smaller institutions

The planned second round would broaden eligibility beyond the largest financial companies to more institutions outside banking and electronic financial service providers. Its detailed measures were confirmed after the fifth meeting of the commission's Frontier AI Situation Response Task Force, involving the Financial Supervisory Service and the Financial Security Institute.

The eligible applicant pool was set to increase from 49 companies to 75, a rise of 26, or about 53%. Available places would rise from ten to as many as 15, an increase of 50%. These figures describe eligibility and planned selection capacity, not the number of applications actually received.

For financial companies, the minimum asset requirement would fall from 10 trillion won to 2 trillion won, while the staffing threshold would fall from 1,000 to 300 permanent employees. Electronic financial service providers would need annual electronic financial transaction amounts of at least 2 trillion won and related revenue exceeding 10% of total revenue.

In both categories, the chief information security officer could not simultaneously hold other information technology functions. Selected companies were to receive a letter of non-enforcement for network separation rules lasting one year, after putting alternative controls in place.

The inspection findings make implementation capacity central to that expansion. A lower size threshold opens participation to more firms, but does not remove the need to demonstrate functioning safeguards. Regulators had also planned to consider further rounds and a permanent framework based on test progress and demand; the postponement did not settle those decisions.

Why Korea is reconsidering physical separation

The policy tension is documented in the Korea Capital Market Institute's network separation report, issue report 24-26, published on November 22, 2024, by researcher Hwang Se-woon. It examines the security benefits and operating costs of Korea's approach, alongside the growing use of cloud software and generative AI.

The report traces the policy's origins to banking disruptions on March 20, 2013, when malicious code disguised as patches spread through antivirus update servers. It says authorities announced comprehensive financial security measures in July 2013 and that network separation became formally established in regulations in 2015.

The scale of digital finance helps explain that response. Electronic transactions already accounted for 87.7% of all financial transactions in 2013, while average daily internet and mobile banking transaction amounts exceeded 33 trillion won, according to the report.

Physical separation uses distinct equipment or infrastructure to keep internet activity apart from internal work. Logical separation uses technology such as virtualization to maintain separate environments on shared physical infrastructure. Network segmentation divides systems into zones and controls traffic between them, rather than necessarily prohibiting all connections.

Hwang's report identifies a practical conflict: much modern software, including generative AI, is delivered through external cloud services, while important financial data sits on protected internal networks. It describes how strict separation can obstruct research, software development and routine use of subscription software, even where such tools could reduce costs or improve work processes.

International comparisons favour controlled connections

The Korea Capital Market Institute report compares approaches in the United States, Europe and Australia. Its central distinction is between prescribing physical disconnection and allowing institutions to choose security designs while holding them responsible for protecting information.

In the United States, the Federal Financial Institutions Examination Council's information security handbook describes trusted and untrusted network zones, controls between those zones, accurate network and data flow diagrams, and layered restrictions. The report finds that this approach does not prescribe one universal physical separation method.

For Europe, the report examines the General Data Protection Regulation, the second Payment Services Directive and European Banking Authority guidance. The guidance it describes calls for management responsibility, controls on external technology providers, limited access rights, and periodic monitoring and testing of access controls.

Its Australian comparison draws on Australian Prudential Regulation Authority guidance and Australian Cyber Security Centre advice. These approaches treat segmentation as one of several security tools and warn against treating an entire internal network as uniformly trusted once an attacker crosses its outer boundary.

The report also cites surveys showing that, in 2022, 82% of surveyed banks had moved or planned to move at least half of their mainframes to the cloud. A separate 2023 survey found that 66% of global financial firms were pursuing cloud migration for growth and innovation. These are different survey measures, not a trend showing adoption falling between years.

The comparison offers a specific lesson for the Korean inspections: greater choice over network design depends on access restrictions and testing actually working. The report recommends gradual reform, standardized data classification, stronger accountability for incidents, security investment and regular supervisory inspections with corrective action.

A separate SaaS exemption keeps data restrictions

A separate easing measure allows financial companies and electronic financial service providers to use certain administrative SaaS tools without entering the financial regulatory sandbox. That route should not be confused with the temporary AI security testing exemptions granted to the ten inspected firms.

The exemption covers SaaS specified under the Enforcement Decree of the Act on the Development of Cloud Computing and Protection of Its Users, in relation to network separation requirements under the Electronic Financial Transactions Act and associated supervisory regulations.

It excludes activities involving personal identification information or personal credit information. SaaS use involving pseudonymised personal data still requires sandbox approval. Pseudonymisation replaces identifying details with substitutes, but does not necessarily make the information anonymous.

Applications must undergo screening by the Financial Security Institute. Firms must maintain security controls for access devices and review compliance controls every six months, reporting the findings to their chief information security officers. The institute also published a booklet explaining financial technology security, SaaS risks and measures to address them.

Taken together, the two routes show that regulatory easing is conditional rather than unrestricted. The temporary testing programme depends on substitute controls before connection; the administrative SaaS route retains data exclusions, prior screening and recurring reviews.

The immediate unresolved issues are which firms had each inspection deficiency, how those weaknesses developed, and when the postponed selection will resume. The institute's confirmation that corrections preceded connections is the strongest stated limit on the incident. The broader test is whether firms can maintain working controls after inspection, rather than producing plans that require inspectors to prompt implementation.

Key Points

• Ten financial firms received temporary network separation exceptions on June 19 for AI and security SaaS testing.
• Inspections from mid-July to early September found many required safeguards still at the planning stage.
• Some security management consoles were accessible from the internet, and a cloud service lacked security certification.
• The Financial Security Institute said all deficiencies were corrected before internal and external networks were connected.
• The October 7 selection of up to 15 more institutions was postponed amid separate breaches and data leaks.
• The planned expansion would increase the eligible applicant pool from 49 to 75 companies.
• A separate administrative SaaS exemption retains restrictions on personal data, institute screening and security reviews every six months.