Your Attacker's AI Knows You Better Than Yours Does

Founder and CEO at CYFIRMA, overseeing the business strategy, product roadmap, growth & expansion plan, and investor relationship.

gettyOn October 9, 2026, Japan declared its first-ever cybersecurity state of emergency. A day earlier, ransomware had spread through IDC Frontier, a SoftBank-owned cloud provider, reaching almost 500 client companies and local governments within hours. Across the water, an open-source AI agent was working through seven South Korean banks and lenders, and President Lee Jae Myung told his cabinet there were signs AI had been used.

Since then, nearly every conversation I have opens with the same question: “Can’t AI just stop this?” In more than two decades in this field, I’ve rarely seen a technology arrive with so much promise.​

For years, Japan and South Korea benefited from two defenses that never appeared on an architecture diagram.

The first was language. Badly translated phishing was easy to spot, and attackers who couldn’t follow Japanese or Korean business conventions looked elsewhere. That advantage is gone, as open-weight models now produce fluent Japanese and Korean.

The second was skill. In Japan, police arrested three students between the ages of 14 and 16 in 2025 who were accused of using ChatGPT to hijack more than 1,000 mobile lines. Police also arrested a 17-year-old accused of using a ChatGPT-built program to pull 7.25 million membership records from an internet cafe chain.

AI hasn’t produced a generation of elite hackers. Rather, it’s created millions of good-enough attackers and given each the reach of a team.​

​The Korean bank campaign used credential stuffing, one of the oldest techniques in the book, run by ARTEX, an autonomous penetration-testing agent published openly on GitHub. It moved through hundreds of IP addresses and altered its timing as it went.

The Japanese wave showed the same pattern. One actor moved through hundreds of addresses in under an hour, shaping each request around the previous response. It found nothing exotic: broken authorization, exposed admin APIs, keys left in mobile apps and flaws that already had patches.

The AI was the accelerant. The fuel was ours.

If your strategy is to out-model the adversary, then you need smarter models. Your model has guardrails, procurement cycles and a vendor who can switch it off. The attacker’s has none of those. The best open-weight models already match closed frontier models at finding vulnerabilities at a fraction of the cost.

Smarter models matter, but a smarter model pointed at the wrong problem just reaches the wrong answer faster.

No single model can protect every industry because each faces distinct attacks. Brokerages are targeted for account takeovers, railways for payment-data theft and operational disruption, and city governments for compromises of the cloud providers they rely on but don’t control. Attackers, entry points and what “normal” looks like all differ.

A generic model trained on broad telemetry erases those distinctions, and that’s where attackers hide. The threat model comes first; the detection model follows.

Vendors love to talk about their models. Few talk about what the model can see. Most organizations feed AI internal logs, which describe what has already happened. In both countries, the agents kept activity from each address low enough that internal detection saw little worth flagging. If your data begins at the perimeter, so does your AI.

What matters is outside-in data—what’s being discussed, sold, scanned and staged about you before the first packet arrives.

Every category is growing, but AI-enabled threats include AI-assisted phishing and reconnaissance, vulnerability discovery, malware, semi-autonomous operations and autonomous ransomware. North Korea’s Kimsuky group used ChatGPT to help forge a South Korean military ID. In November 2025, Anthropic disclosed a campaign in which AI carried out an estimated 80% to 90% of the tactical work.

The discipline is to pick the vectors that matter to your sector, geography and exposure, and then build depth there.

Attackers think in paths; defenders too often think in lists. A list tells you that you have 4,000 vulnerabilities. A path tells you that a mobile app leaks a key, the key opens an admin API and the API can export your customer table. The only fair response is to map the routes first and break the earliest, cheapest link.

AI is remarkably good at this if it has a reliable asset inventory, a third-party map, the business logic and a view of who’s interested in you. Without that context, it’s guessing.

Is AI enough on its own to defend against cyberattacks? No, and I say that as someone who has built a company around AI-driven intelligence. AI is now necessary because machine-speed adversaries can’t be met at human speed, but it’s a multiplier, and multiplying zero still gives you zero.

The incidents in Japan and South Korea showed that the attacker’s advantage was never the model. What remained was everything we had failed to fix, and machines found it in hours. A smarter model won’t change that. Knowing your industry’s real threats, the paths to what matters most and what the outside world already knows about you will.

The question for every CXO and board is: What does an attacker’s AI already know about you that yours doesn’t? That’s what an external threat assessment is for: identifying the exposed assets, leaked credentials, contractor weaknesses and industry-specific paths that machines can now find and exploit in hours and turning them into the outside-in data, vectors and attack paths your AI should be watching.

The attackers in Japan and South Korea won because their AI knew us better than ours did. Until that changes, a smarter model just gives us a faster way to be surprised.

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?