AI didn't make hackers smarter. It made hygiene non-negotiable

What does a super-hacker look like in 2026? If the latest case out of South Korea is any guide, he's allegedly 26, lives in southern China, and left enough traces online that investigators could work out his name, age and location.

According to a Washington Post report republished by the Anchorage Daily News, CrowdStrike alleges that this man, an amateur who used a few AI chatbots, was behind attacks on at least nine major South Korean banks. Tens of thousands of South Koreans had financial data exposed. CrowdStrike also found that he asked a chatbot to help sell the stolen data and to draft a fake résumé touting cybersecurity skills. CrowdStrike's Adam Meyers put it bluntly: "Anyone can be a super-hacker now."

The same report says the NSA advised websites to shut down if they couldn't immediately patch a major PHP hole discovered with AI tools. It also says Anthropic held back its Mythos model because it was too good at hacking, and the US government has restricted who gets the top models without filters.

The new threat isn't the genius, it's the crowd

The cinematic version of AI hacking is a superintelligent adversary. The documented version is duller and, we think, scarier: a mediocre operator with expert-grade tooling. A SingularityHub summary of research describes a group of relative amateurs that "achieved an operational scale that would have previously required a significantly larger and more skilled team." That attack was largely unsuccessful, but the scale is the point.

This changes defender math. Security planning has long relied on an implicit filter: most attackers lack the skill to exploit most weaknesses, so you could be imperfect and still be fine. In our view, that filter is thinning. Every exposed, unpatched, misconfigured system now faces a bigger pool of people able to find it.

Basics matter more because the attacker's patience is now free

Here's where we'd push back on the doom narrative. The South Korean attacks were devastating, but the weakness documented on the attacker's side was his own sloppiness: traces online and unsecured systems let investigators identify him. For defenders, the equivalent lesson is that AI rewards whoever is faster and more consistent, and basics are exactly what consistency looks like: patching quickly, limiting privileges, knowing what you have exposed.

Anthropic's Jason Clinton warned that things "are going to get a lot worse before they get better." The NSA's PHP advice is the clearest illustration: when a flaw affects a language used by an estimated 70 percent of websites, the window between disclosure and exploitation is the whole game. If your patching process takes weeks, we think you are handing AI-assisted attackers a head start.

Oracle offers a glimpse of the defensive version. Its customer success chief Gary Miller told Gadget the company uses one model to find a vulnerability, another to check that it's real, and then models to code and test the fix. He added that some vulnerabilities humans had found were not being found by the models, so experts now give them hints. That's a sober, useful corrective: AI helps, but it doesn't replace people who know the codebase.

The access gap is the real policy problem

The most contestable part of the story is the lockdown on top-tier models. The Washington Post reports that critics say more defenders need access, and Vercel's CTO Malte Ubl argued the delay "cost the industry at least three, four, five months." Meanwhile, open-weight models, mostly from Chinese companies, perform at close to the same level as Mythos did at its April debut, according to Anthropic.

Our view: restricting capabilities that are about to be widely available anyway mostly burdens the honest. Attackers will use whatever models and tools they can, as Anthropic's Logan Graham put it. A small business is unlikely to land on a government-approved access list, but it can still turn on multifactor authentication and automate updates. Policy should make defensive AI easier to get, not harder.

What to do on Monday

Treat every internet-facing system as if a competent adversary is already probing it, because, in our view, the cost of being probed has collapsed. Shorten your patch cycle. Strip permissions down. Keep logs. And if you deploy AI agents yourself, give them the same least-privilege treatment: ITWire's guidance on open source agents stresses limiting what they can access and requiring human approval for high-impact changes.

The reckoning isn't a mysterious new kind of attack. It's the end of being safe by obscurity and by attacker laziness. The organizations that survive it won't be the ones with the fanciest AI. They'll be the ones that finally did the boring work.