Your Attacker's AI Knows You Better Than Yours Does
In October 2026, Japan declared its first-ever cybersecurity state of emergency after ransomware spread through SoftBank's cloud provider, affecting nearly 500 companies and governments, while AI-powered attacks simultaneously targeted seven South Korean banks.
Since these incidents, cybersecurity professionals have repeatedly asked whether AI can stop these attacks. Despite the technology's promise, recent events reveal a troubling gap between defensive and offensive capabilities. The attacks exploited fundamental vulnerabilities that organizations had failed to address, with AI acting as an accelerant rather than the root cause.
Eroded Defenses
Japan and South Korea had long benefited from two key defenses that weren't reflected in their technical architecture: language barriers and skill gaps. Poorly translated phishing attempts were easily recognizable, and attackers unfamiliar with local business conventions struggled. However, open-weight AI models now produce fluent Japanese and Korean, eliminating this advantage.
The skill gap has also closed dramatically. In 2025, Japanese police arrested three students aged 14-16 accused of using ChatGPT to hijack over 1,000 mobile lines, and a 17-year-old who allegedly used ChatGPT to steal 7.25 million membership records from an internet cafe chain. AI hasn't created elite hackers but has enabled millions of 'good-enough' attackers with the reach of entire teams.
Attack Patterns
The Korean bank campaign utilized credential stuffing through ARTEX, an autonomous penetration-testing agent published on GitHub. It moved through hundreds of IP addresses, altering its timing to avoid detection. Similarly, the Japanese ransomware attack involved one actor moving rapidly through numerous addresses, exploiting basic vulnerabilities like broken authorization, exposed admin APIs, and unpatched flaws.
The AI was the accelerant. The fuel was ours.
Defensive Challenges
Defensive strategies focusing on out-modeling adversaries face significant obstacles. Attackers' models lack guardrails, procurement cycles, and vendor controls that defensive systems require. The best open-weight models already match closed frontier models in finding vulnerabilities at a fraction of the cost.
Generic detection models trained on broad telemetry erase the distinctions between industries, allowing attackers to hide. Different sectors face unique threats—brokerages for account takeovers, railways for payment data theft, city governments for cloud provider compromises. Effective defense requires starting with the threat model, not the detection model.
Most organizations feed AI internal logs describing what has already happened, but the attackers kept activity low enough to avoid detection. Effective defense requires outside-in data—information about what's being discussed, sold, scanned, and staged before any attack begins.
Future Threats
AI-enabled threats are growing across multiple vectors, including phishing, reconnaissance, vulnerability discovery, malware, semi-autonomous operations, and autonomous ransomware. North Korea's Kimsuky group used ChatGPT to forge South Korean military IDs, while Anthropic disclosed a campaign where AI performed 80-90% of tactical work.
The key to defense lies in understanding attackers think in paths, not lists. While a list might identify 4,000 vulnerabilities, a path reveals how a mobile app leak could lead to an admin API that exports customer data. Mapping these routes and breaking the earliest, cheapest link is essential. AI can help with this if provided with reliable asset inventories, third-party maps, business logic, and visibility into who's targeting an organization.
Related
