AI Is Cutting the Time to Find a Hack From Weeks to Hours. Here Is What Security Teams Should Do About It.

Key points

• Park Chan-am, CEO of security firm Steelion and a former adviser to South Korea's National Police Agency, told a government audience in 2026 that AI has cut the time needed to find exploitable software flaws from four weeks to under one day.
• Over 300,000 servers running local AI testing software were confirmed exposed to the open internet in 2026, up from roughly 1,000 when the same vulnerability was first flagged in 2024.
• Park argued that traditional vulnerability scoring systems are no longer enough, and that security teams should instead prioritise flaws already being used in real criminal attacks.
• Park warned that AI agents, software programs that carry out tasks automatically on behalf of users, can leak sensitive data if given too many system permissions.
• He concluded that no technology replaces human judgement in confirming a real threat.

Park Chan-am was breaking into computer systems competitively at age eleven. He went on to advise South Korean government agencies, helped defend the country against cyberattacks linked to North Korea, and now runs a security company called Steelion. This month, as CSO Online reported, he addressed a forum organised around South Korea's fifteenth national Information Security Day, and his message was blunt: AI has changed the rules, and most organisations have not caught up.

How much faster can criminals actually work now?

Significantly faster. Park put a number on it. Finding a usable weakness in a company's software systems used to take a skilled criminal at least four weeks. With AI tools, the same job now takes less than a day. That single shift matters because it means a company that once had weeks to notice a problem and fix it now may have hours.

This changes what counts as a target. Every piece of software a business runs, its accounting tools, its email system, its internal databases, is now a more attractive entry point than it used to be.

Park also flagged a rising blind spot: local AI testing. When staff experiment with open-source AI models on their work laptops or office servers, they often leave those servers open to the internet by accident. A 2024 security flaw in Ollama, a popular tool for running AI models locally, exposed roughly 1,000 servers when it was discovered. By 2026, the same category of exposed servers had grown to over 300,000. Testing AI, Park warned, is itself a security risk.

For companies using AI agents, software programs that read emails, summarise documents, and carry out other tasks automatically, Park raised a specific concern about permissions. If an AI agent has access to more systems than it needs, a criminal who tricks it with a poisoned document can use that access to pull out confidential data. The fix, Park said, is simple in principle but often skipped: design what the AI is allowed to touch before deciding what it will do.

On prioritising which software flaws to fix first, Park pushed back against the Common Vulnerability Scoring System (CVSS), a standard scale security teams use to rank how dangerous a flaw is. A CVSS score alone, he argued, tells you how bad a flaw looks on paper. It does not tell you whether real criminals are actively using it right now. He suggested pairing that score with the Exploit Prediction Scoring System (EPSS), which estimates the real-world probability of a flaw being attacked, and the US government's Known Exploited Vulnerabilities (KEV) list, a catalogue of flaws already weaponised in actual incidents.

A flaw with a middling CVSS score of 7.5 looks unurgent. The same flaw, if it sits on the KEV list and its EPSS probability has jumped from one percent to ninety percent in two months, is on fire. Patch it first.

If your organisation runs AI tools or stores customer data, ask your IT team two questions: which of our servers are visible to the internet that should not be, and which unpatched flaws appear on the US KEV list? Those are the places to start.