Cyber Summit Korea (CSK) 2026. (Source: CSK 2026 Cyber Summit Korea website)

CSK 2026 Session A-1

Current Response to High-Performance AI Threats: State and Challenges

As artificial intelligence (AI) is put to use across cybersecurity ever more rapidly, the way both attacks and defense work is changing too.

In the past, attackers had to find vulnerabilities and prepare attack tools themselves; now, the environment is evolving to where AI can automate parts of vulnerability discovery, analysis, and the attack process.

On the flip side, defenders too are moving toward using AI to analyze huge numbers of security events, spot anomalies, and automate parts of the response process.

Against this backdrop, Cyber Summit Korea (CSK) 2026 featured a session dedicated to the new cyber threats posed by high-performance AI and the current state of response to them.

In particular, Session A-1, "Current Response to High-Performance AI Threats: State and Challenges," held on September 17, 2026, focused on autonomous AI attacks, AI-based security operations, and vulnerability discovery using LLMs, discussing the challenges today's security response frameworks face.

1. Why did CSK 2026 address 'high-performance AI threats'?

CSK 2026 is a global cybersecurity event jointly hosted by the National Intelligence Service and the National Security Research Institute.

The theme of the 2026 event was set as "Staying Ahead for Global Cybersecurity: A Connected World, Security Together."

The event drew participation from foreign intelligence and security agencies, government, industry, and academia, with AI threat response and cybersecurity strategy for the age of AI transformation among the major agenda items.

In its opening remarks, the National Intelligence Service emphasized that a hyper-connected environment is both the foundation for innovation and a pathway cyber threats can travel through, and that the emergence of high-performance AI in particular is raising the level of threat.

What matters here isn't simply the observation that"AI can be used for hacking."

It's that as AI becomes able to carry out more and more tasks on its own, the speed and methods of existing security responses themselves need to be reexamined.

2. What was discussed in Session A-1?

According to CSK's official program, Session A-1 ran from 1:30 to 3:00 p.m. on September 17, 2026, chaired by Professor Kwon Tae-kyoung of Seoul National University.

The session featured three main presentations.

Main presentations

① Threats from AI autonomous attacks and how to respond

Professor Lee Sang-geun, Korea University

② Agentic SOC strategy for countering machine-speed threats

Choi Jae-ho, Executive Director, SK Shieldus

③ Discovering and executing vulnerabilities using LLMs

Professor Lee Byung-young, Seoul National University

This was followed by an expert panel discussing the current level of AI threat response, the direction of future development, and policy responses.

In other words, this session went beyond a simple introduction to AI technology, and can be viewed around questions like:

How much faster is AI making attacks?

Can defenders keep pace with that speed?

Are existing security systems still adequate in the age of AI?

3. Why AI autonomous attacks matter

In traditional cyberattacks, attackers often had to carry out several stages themselves.

For example,

gathering information → finding vulnerabilities → deciding on an attack method → executing the attack → analyzing the results

— this kind of process was required.

When AI gets involved in this process, it can automate part of the work.

For instance, AI can analyze large volumes of data, identify characteristics of code or a system, and suggest the next step based on what it finds.

This is why the security industry views AI's potential to raise the level of automation in the attack process as an important issue.

CSK 2026 addressed this under the theme of 'AI autonomous attacks.'

That said, this shouldn't immediately be taken to mean"AI carries out every hack completely on its own."

Today's AI systems can still make errors depending on the environment, and controlling the permissions and tools they're connected to matters too.

So the key point isn't the simple claim that AI fully replaces the attacker, but rather

that the speed and level of automation of the many tasks an attack requires can increase

4. In the 'machine-speed' era, defense has to speed up too

One important characteristic of AI-driven attacks is speed.

If AI can handle at scale tasks that a human would otherwise have to check one at a time, defenders may find it hard to keep responding the way they always have.

The Agentic SOC (Security Operations Center) introduced in CSK 2026's Session A-1 connects to exactly this issue.

Where a traditional SOC centers on security operators reviewing huge numbers of events, analyzing anomalies, and then responding, an Agentic SOC is an approach that uses AI agents and similar tools to expand automation across detection, analysis, and response.

Put simply, it looks like this.

Traditional security operations

Enormous volume of logs

↓

Security system detects

↓

Operator analyzes

↓

Response method decided

↓

Action taken

The direction of AI-based security operations

Enormous volume of logs

↓

AI analyzes in real time

↓

Supports threat assessment

↓

Automated response or operator approval

↓

Action taken and results analyzed

What matters here isn't unconditionally handing AI every permission.

The key is letting AI analyze and respond quickly while still putting permission limits and verification steps around critical tasks.

5. LLMs can also be used to find vulnerabilities

The third presentation topic was "Discovering and Executing Vulnerabilities Using LLMs."

LLMs can be used not only for natural language processing, but also for analyzing and generating program code.

This is why research is underway in the security field into using LLMs to analyze a piece of software's code structure or find potential security issues.

Conversely, attackers too could use AI for code analysis and attack preparation.

The fact that CSK 2026's session lineup treated this dual-use nature as its own separate topic is related to this shift.

6. AI security doesn't end at protecting just the 'model'

When people talk about AI security, it's easy to think only of the AI model itself.

But in real services, AI connects to a range of systems.

For example,

User

↓

AI application

↓

LLM

↓

Internal company data

↓

Retrieval system (RAG)

↓

External APIs

↓

Business systems

— a structure like this can take shape.

When AI is connected to external systems this way, what matters isn't just the security of the model itself, but also access management, data protection, API security, the supply chain, and output validation.

OWASP's LLM security guide also treats prompt injection, sensitive information disclosure, supply chain vulnerabilities, data and model poisoning, improper output handling, and excessive agency as major risks.

In particular, as AI agents increasingly get to use external tools directly, what permissions to grant AI is becoming an even more critical security issue.

7. In the age of AI agents, 'permissions' are the key

One of the biggest differences between an AI chatbot and an AI agent is the scope of action.

A simple chatbot is centered on answering questions.

An AI agent, on the other hand, can use external tools, query data, and carry out specific tasks depending on how it's configured.

So if an AI agent makes a bad judgment or is influenced by malicious input, the impact can go beyond a wrong answer and actually affect real systems.

For example, imagine a company's AI agent is connected to the following.

Internal documents
Email
Databases
Cloud services
Business APIs
File storage

In that case, the core question in AI security isn't simply

"Is the AI safe?"

but rather

"What can the AI access, and what actions can it actually take?"

8. Why existing security approaches alone may not be enough

AI technology is advancing fast.

So an approach of building a security system once and running it unchanged for a long time can start to hit its limits.

In particular, if AI attack methods keep changing, the security system needs to be continuously re-verified as well.

The LLM security guide OWASP published in 2026 reflects real AI security incidents and research findings, expanding the scope of risk and emphasizing managing risk across the entire development, deployment, and operation lifecycle of AI applications.

NIST, too, is developing a Cyber AI Profile to separately manage cybersecurity risk for AI systems. NIST's initial documentation lays out an approach for identifying, prioritizing, and managing the cybersecurity risks that come with adopting AI.

In the end, AI security is becoming less about installing a single security product, and more about a process of continuously checking AI's entire lifecycle.

9. Which is why 'AI red teaming' matters

The method of testing a system from an attacker's perspective to find vulnerabilities in an AI system is called red teaming.

Put simply, it means

testing in advance what could go wrong if you imagine attacking this AI system

As AI keeps advancing, security staff need to keep testing AI systems continuously as well.

In fact, the National Intelligence Service disclosed plans to expand its LLM-centered AI safety evaluation to cover AI agents as well, and to prepare AI red-teaming guidelines and a security playbook so each institution can check its own AI systems for vulnerabilities.

This approach represents an important shift in AI security.

10. Why a digital twin is needed

Continuously running attack tests against a live, in-production system using AI isn't realistically easy.

That's because a mishandled attack test on a real system can cause a service outage or affect important data.

So one approach is to build a separate environment similar to the real system and test attack scenarios there.

To put it simply, it looks like this.

The real system

Corporate network

→ servers

→ databases

→ business systems

→ AI systems

The test environment

Digital twin

→ servers similar to the real ones

→ data structures similar to the real ones

→ an AI environment similar to the real one

→ attack/defense testing

In this kind of environment, an AI red team can repeatedly test a range of attack scenarios and identify vulnerabilities.

Coverage of CSK 2026's session also discussed a continuous red-teaming approach using a digital twin environment similar to the real system as one direction for response.

11. Another shift visible at CSK 2026

CSK 2026 also featured the international cyber exercise APEX 2026 alongside its conference sessions.

According to official materials, 23 countries took part in APEX 2026, and this year in particular saw the new formation of an AI-driven Blue Team.

The goal was to verify, in a real exercise environment, the latest technology for countering AI-driven threats and autonomous defense capability.

This is a very important trend.

That's because if AI is increasingly likely to be used in cyberattacks, defense needs to make use of AI too.

In other words, a new competitive structure is taking shape:

AI attack

↕

AI defense

12. How will AI security change going forward?

Bringing together what was discussed in CSK 2026's Session A-1, the following shifts are likely to matter for AI security going forward.

① Human-centered security → human + AI security

Rather than security staff handling every event directly, the model where AI first analyzes large volumes of information and humans handle the critical decisions is likely to expand.

② Periodic checks → continuous verification

Because AI attack methods can change quickly, checking a few times a year may not be enough anymore.

A structure for continuously evaluating and testing the AI system itself becomes important.

③ Model security → AI ecosystem security

It's not enough to protect just the AI model — data, applications, APIs, agents, external tools, and the cloud environment all need protection together.

④ Ordinary SOC → Agentic SOC

Using AI agents to automate the analysis of and response to security events is likely to expand.

⑤ Attack simulation → AI red teaming

The process of continuously testing an AI system from an attacker's perspective and fixing the vulnerabilities found can become important.

13. AI security everyday users should know about

This isn't a story relevant only to companies or national agencies.

Everyday users, too, should keep a few basic principles in mind when using generative AI.

Don't enter sensitive personal information directly

Before entering sensitive information like ID numbers, passwords, or financial details into an AI service, check that service's data handling policy.

Don't immediately run links or files AI gives you

Just because an answer was generated by AI doesn't mean it's always safe.

It's especially important to get in the habit of separately verifying external links, programs, scripts, and files.

Don't treat AI's answer as the sole basis for a security decision

AI can generate information that sounds plausible but is wrong.

Important security-related decisions need to be backed by official documentation or reviewed by security staff.

Don't grant AI excessive permissions

When using an AI agent, it's important to grant access to files, services, and APIs only within the scope that's actually needed.

14. The core message of CSK 2026's Session A-1

CSK 2026's "Current Response to High-Performance AI Threats: State and Challenges" session doesn't view AI-era cybersecurity simply through the lens of "AI can be used for hacking."

The key point is that both attack and defense are speeding up.

AI can be used to analyze vulnerabilities and automate security work, but if exploited by an attacker, it can also speed up parts of the attack preparation and execution process.

So the following question becomes important for the cybersecurity of the future.

Can AI-driven defense keep pace with the speed of AI-driven attacks?

And beyond that,

how much autonomy and authority can be safely given to AI?

is a question that matters just as much.

The fact that AI autonomous attacks, Agentic SOC, LLM-based vulnerability analysis, and AI red teaming were all discussed together at CSK 2026 connects to this same shift.

15. Conclusion

AI is a technology that both creates new risks for cybersecurity and can strengthen defensive capability at the same time.

So rather than viewing AI itself as an unconditionally dangerous technology, what matters is examining in what environment AI has what permissions, and what actions it can actually take.

In particular, in an environment where AI agents are connected to corporate systems and automatically carry out multiple tasks, existing security approaches alone may struggle to manage every risk.

Going forward, a connected security framework of

AI security evaluation → continuous red teaming → AI-based detection → permission control → automated response → final human verification

is expected to become increasingly important.

CSK 2026's Session A-1 can be seen as an example that let us examine exactly this question — how cyber defense needs to change in the age of AI.

Please refer to the Cyber Summit Korea (CSK) 2026 website (https://cybersummit.kr)