White hat hacker Chanam Park zeros in on the AI era’s key security challenges

Dubbed the “Genius Hacker,” Chanam Park began his white hat hacker journey at the precocious age of 11, winning awards at domestic and international hacking competitions since his teenage years.

He has since served as a cybersecurity advisor for various Korean government agencies, including the National Police Agency, and has played a key role in the country’s defense against Democratic People’s Republic of Korea (DPRK)-affiliated cyberattacks.

At a seminar held this month as part of the 15th Information Security Day event hosted and organized by government agencies including the Ministry of Science and ICT and the Korea Internet and Security Agency (KISA), Park, now CEO of security firm Steelion, explained the changes in the security environment in the AI ​​era and the priority response tasks for information security organizations under the theme of “Major AI Threats and Security Priorities.”

“While AI is a new technology, the core of security ultimately lies in access control, supply chain management, and human verification,” he said.

Like many security experts, Park sees AI fundamentally changing the speed of cyberattacks. In the past, infiltrating a corporate network required significant time analyzing a range of software systems to find exploitable vulnerabilities — a task that the use of AI has significantly accelerated.

“In the past, it took at least four weeks to find vulnerabilities, but now it takes less than a day,” he said. “In the era of AI, all software installed within a company becomes a much more critical target for attacks.”

As a result, securing internal software and the software supply chain are paramount — and require a different perspective on accountability, Park noted.

“Clients often ask, ‘Isn’t this just a product made by the vendor?’” he said. “From the moment it is installed in the company system, that software is no longer a vendor issue but part of the corporate system.”

“We are now in an era where third-party issues can no longer be attributed solely to vendor responsibility,” he stressed.

MCP under threat

Park also sees authorization management as a key challenge security teams will face in the agentic era. For example, companies have been increasingly utilizing Model Context Protocol (MCP)-based AI agents to read emails, analyze documents, and connect internal systems with various business tasks. But as the workload handled by AI agents increases, every step an AI agent takes, reading external documents and interacting with internal systems, can become a potential attack vector.

Prompt contamination through malicious documents and the leakage of internal information via agents with excessive privileges are quite realistic scenarios, Park said. In particular, he pointed out that issues that previously ended as minor problems, such as residual privileges left by former employees or outsourced personnel, could escalate into major incidents as AI automatically links these elements together.

“When introducing AI, permissions must be designed before functions,” he said. “The entire MCP process must be approached as a single attack path.”

The ever-widening blast radius of AI testing

Local AI testing environments are becoming a dangerous security blind spot that information security leaders often overlook. Rapid experimentation with open-source AI, such as LLaMA-based models, on personal or work PCs often results in servers or ports being left open, and if vulnerabilities are discovered, intrusion pathways immediately open up.

“When the [Ollama] remote code execution vulnerability was discovered in 2024, there were over 1,000 servers exposed to the internet, but recently in 2026, it has been confirmed that over 300,000 servers from the same targets are exposed,” said Park, adding that “the act of testing AI itself can become a new security risk.”

Vulnerability management on notice

Security operations must also change, Park stressed, noting that the number of alerts that security personnel must handle has increased tenfold, and in some cases up to a hundredfold, making it virtually impossible to respond to all vulnerabilities using the same standards.

As a solution, Park sees the Common Vulnerability Scoring System (CVSS) being insufficient for determining priorities. Instead, he suggested that vulnerability response priorities be determined by utilizing the Exploit Prediction Scoring System (EPSS), which predicts the actual likelihood of exploitation, along with the US government’s Known Exploited Vulnerabilities (KEV) list.

For example, if a vulnerability’s CVSS score is 7.5, it is not classified as critical, so it is likely to be pushed down the priority list. But the response priority changes completely if the same vulnerability is listed on the KEV list, has been exploited in actual ransomware attacks, and the probability of an attack based on EPSS has skyrocketed from 1% to 90% within two months. “You must consider these factors together to identify the vulnerabilities that actually need to be patched first,” he stressed.

“Amidst the vast noise known as the AI s​lop, the criteria for deciding what to patch first is now becoming a core competency for security personnel,” he added.

Park also presented new defense techniques applicable to the AI ​​era, such as methods to detect automated attacks by analyzing behavioral differences between humans and AI attackers, and proof of work (PoW) challenges that intentionally impose computational load on AI attackers to slow down their attacks. A prime example is filtering out abnormal access by analyzing mouse movements, keyboard input, and scrolling patterns.

“It is a more realistic strategy to reduce the burden on security teams by filtering out at least some attacks, rather than trying to block them 100%,” he said.

Even in the age of AI, technology alone cannot ensure complete security, he noted. “While AI can scan for threats broadly and quickly, verifying and confirming them ultimately falls to humans,” he said. “Humans are still important.”