Cisco's Relentless Defense Report, based on a survey of 8,000 security professionals across 30 countries, found that Korean companies' AI security capability falls short of the global average. On a 100-point scale, the average cyber defense score was 64 globally and 58 in Korea. Only 8% of organizations worldwide and 1% in Korea had reached the 'Relentless Defender' level, meaning they continuously defend and improve in the AI era. The share of organizations confident they could detect and respond when an AI agent is compromised was 54% globally and 37% in Korea.

Sung-kyu Hwang, Senior Vice President for Security at Cisco Korea, identified organizational issues, rather than technology, as the biggest obstacle to security. Six in ten practitioners who struggle with security response named organizational capabilities such as collaboration, approval processes and data accessibility as major areas for improvement. Forty-one percent of respondents globally and 24% in Korea said their security, networking and IT teams share visibility and tools and work as one team. In Korea, 40% said security metrics are too technical to persuade executives, and 44% said they spend more time collecting and connecting data scattered across systems than responding to actual threats.

Top defenders were ahead on organizational capability. Of these, 83% operated integrated security teams that share information and tools across security, network and infrastructure, and 86% applied consistent security controls to AI agents and non-human identities. The remaining group scored 37% on both measures. Among top defenders, 71% said cyber incidents declined after they expanded security investment. Hwang noted that top defenders resolve collaboration and persuasion issues through governance first, rather than simply increasing budgets. More than 8 in 10 security practitioners said AI use had improved threat detection and response capability by at least 25%, a figure Korea matched. Cisco proposed five action items: clarifying authority and responsibility, integrating separated teams, repeating hands-on training, actively using AI, and choosing paths that members can easily follow.

Mukesh Chand Taplyal, Cisco's SASE business leader for APJC, said AI agents are changing traffic patterns. He explained that an agent performing a single task generates 450% more traffic than a human. He noted that humans have a limited scope of access and are constrained in speed and scale but have judgment, whereas agents have broad access and scale exponentially but lack judgment and common sense, making them a major organizational risk. His proposed response principles are visibility, identification, authorization and real-time response. To counter shadow AI, data leakage and prompt injection, organizations need visibility into interactions with internal AI applications, and open-source AI models should be checked for malicious code, licensing and provenance before execution. He stressed that each agent must be identified by its owner and purpose, and every action authorized based on the data it accesses, the source of the request and the data transfer path. He said the controllable factor is how quickly organizations respond to change and risk, and adapting to risk in real time should be the strategy.

Tin Satchidev, Cisco's APJC chief security architect, offered guidance for budget-constrained companies. He said every prompt and response must be examined, and static methods that only look for specific data such as credit card numbers are insufficient. He expected model-based inspection to raise costs somewhat in the short term but saw AI as an opportunity to lower security costs over the long term. He advised focusing not on doing something new but on using agentic AI to accelerate security challenges that organizations have struggled with for years. Citing a microsegmentation project, he said work that took five to ten people six months might be completed by AI in five days. He also stressed model selection is key: small language models focused on specific security tasks such as event log analysis, source code vulnerability scanning and supply chain checks are inexpensive to run and can operate on-premises, keeping data from leaving the organization. For budget-limited organizations he recommended combining open source with agents, and said Cisco releases its AI security tools as open source first to build trust and raise security standards across the industry. He advised that organizations that verify effectiveness through open source can later adopt the enterprise-supported version of DefenseClaw (Cisco's open-source security governance tool) once budget is secured.

Cisco's Relentless Defense Report, based on a survey of 8,000 security professionals across 30 countries, found that Korean companies' AI security capability falls short of the global average. On a 100-point scale, the average cyber defense score was 64 globally and 58 in Korea. Only 8% of organizations worldwide and 1% in Korea had reached the 'Relentless Defender' level, meaning they continuously defend and improve in the AI era. The share of organizations confident they could detect and respond when an AI agent is compromised was 54% globally and 37% in Korea.

Sung-kyu Hwang, Senior Vice President for Security at Cisco Korea, identified organizational issues, rather than technology, as the biggest obstacle to security. Six in ten practitioners who struggle with security response named organizational capabilities such as collaboration, approval processes and data accessibility as major areas for improvement. Forty-one percent of respondents globally and 24% in Korea said their security, networking and IT teams share visibility and tools and work as one team. In Korea, 40% said security metrics are too technical to persuade executives, and 44% said they spend more time collecting and connecting data scattered across systems than responding to actual threats.

Top defenders were ahead on organizational capability. Of these, 83% operated integrated security teams that share information and tools across security, network and infrastructure, and 86% applied consistent security controls to AI agents and non-human identities. The remaining group scored 37% on both measures. Among top defenders, 71% said cyber incidents declined after they expanded security investment. Hwang noted that top defenders resolve collaboration and persuasion issues through governance first, rather than simply increasing budgets. More than 8 in 10 security practitioners said AI use had improved threat detection and response capability by at least 25%, a figure Korea matched. Cisco proposed five action items: clarifying authority and responsibility, integrating separated teams, repeating hands-on training, actively using AI, and choosing paths that members can easily follow.

Mukesh Chand Taplyal, Cisco's SASE business leader for APJC, said AI agents are changing traffic patterns. He explained that an agent performing a single task generates 450% more traffic than a human. He noted that humans have a limited scope of access and are constrained in speed and scale but have judgment, whereas agents have broad access and scale exponentially but lack judgment and common sense, making them a major organizational risk. His proposed response principles are visibility, identification, authorization and real-time response. To counter shadow AI, data leakage and prompt injection, organizations need visibility into interactions with internal AI applications, and open-source AI models should be checked for malicious code, licensing and provenance before execution. He stressed that each agent must be identified by its owner and purpose, and every action authorized based on the data it accesses, the source of the request and the data transfer path. He said the controllable factor is how quickly organizations respond to change and risk, and adapting to risk in real time should be the strategy.

Tin Satchidev, Cisco's APJC chief security architect, offered guidance for budget-constrained companies. He said every prompt and response must be examined, and static methods that only look for specific data such as credit card numbers are insufficient. He expected model-based inspection to raise costs somewhat in the short term but saw AI as an opportunity to lower security costs over the long term. He advised focusing not on doing something new but on using agentic AI to accelerate security challenges that organizations have struggled with for years. Citing a microsegmentation project, he said work that took five to ten people six months might be completed by AI in five days. He also stressed model selection is key: small language models focused on specific security tasks such as event log analysis, source code vulnerability scanning and supply chain checks are inexpensive to run and can operate on-premises, keeping data from leaving the organization. For budget-limited organizations he recommended combining open source with agents, and said Cisco releases its AI security tools as open source first to build trust and raise security standards across the industry. He advised that organizations that verify effectiveness through open source can later adopt the enterprise-supported version of DefenseClaw (Cisco's open-source security governance tool) once budget is secured.