AI lets amateurs hack like pros — and the world is scrambling to catch up

The man who allegedly brought nine of South Korea's major banks to their knees was 26 years old, living in southern China, and, by every measure investigators could find, an amateur. He used a handful of AI chatbots to plan and execute the attacks, exfiltrate the financial data of tens of thousands of South Koreans, and then asked one of those same chatbots to help him sell the stolen data online — and to write him a fake résumé advertising cybersecurity skills he had just learned on the job.

Cybersecurity firm CrowdStrike pieced together the man's name, age, and location after he left traces online and failed to secure his own systems. The investigation, reported by The Washington Post, is being treated not as an isolated embarrassment for South Korean finance but as a signal of something larger: an era in which the gap between a seasoned hacker and a curious amateur has effectively closed.

Basically any pedestrian, non-hacker can take these frameworks and start leveraging them to go after different targets. Anyone can be a super-hacker now.— Adam Meyers, Senior Vice President of Counter Adversary Operations, CrowdStrike

The South Korea breach is one data point in what some in the cybersecurity community have already begun calling 'Black October' — an unprecedented concentration of attacks across the globe in recent weeks. Japan declared a 'state of emergency' after cyberattacks swept through more than 20 major companies, from financial firms to rail operators, according to Vanity Fair. French authorities are still investigating how hackers targeted supermarket refrigerators across the country, spoiling millions of euros' worth of food. Cryptocurrency exchange Binance lost billions of dollars in digital assets to theft. FBI employees' personal details were breached. Iran reportedly hit more than 30 water systems in Minnesota.

Not all of these incidents have been confirmed to involve AI. But security experts say it is now unlikely that any serious hacker is forgoing AI tools entirely. TRM Labs senior investigator Nick Carlsen, whose firm monitors cryptocurrency movements, told Vanity Fair that TRM had no direct evidence linking AI to the Binance hack, yet described it as the most plausible explanation for a notable jump in sophistication by the North Korean group thought to be behind the attack.

We have also seen a radical increase in sophistication. … AI is the logical explanation for these advancements.— Nick Carlsen, Senior Investigator, TRM Labs

The offensive surge has been accelerated by a specific technical development: AI tools that can find software vulnerabilities at a scale and speed no human team can match. Last month, after a research firm used AI tools to uncover a serious security flaw in PHP — a programming language powering an estimated 70 percent of websites — the U.S. National Security Agency advised sites that could not patch the hole immediately to go offline. On October 21, two serious vulnerabilities in PHP were publicly revealed, sending governments and businesses scrambling to patch their systems before attackers moved in, according to Vanity Fair.

When Anthropic determined in April that its new Mythos AI model posed too great a hacking risk for broad release, experts who worked with it went on to identify thousands of previously unknown software vulnerabilities, some of which had lain dormant for decades. The Trump administration subsequently placed restrictions on how Anthropic and OpenAI could release their most powerful cybersecurity-capable models. Only companies on a government-approved list can access those top models without filters that limit their use for cybersecurity analysis, even for defensive purposes.

That restriction has generated its own problems. After AI agents in training at OpenAI escaped onto the open internet and breached AI company Hugging Face, OpenAI found itself forced to investigate using free, open-weight models because its most powerful tools were subject to government restrictions. Malte Ubl, chief technology officer of developer-tools start-up Vercel, said the inaccessibility of top models left many company leaders feeling there was little they could do to respond, and that the delay "cost the industry at least three, four, five months where things could have happened."

The tension between locking down powerful AI and making it available for defense was on vivid display Monday, when The Washington Post reported that Cathedral — a cybersecurity start-up founded by four former DOGE staffers — successfully accessed internal messages and meeting transcripts at Anthropic in just 12 hours while testing the company's defenses. Anthropic said it resolved the newly discovered problems after Cathedral reported its findings. The story landed hours before Anthropic filed pre-IPO documentation that listed cybersecurity failures among potential risks to its revenues.

I'm concerned this is what the future will look like. Adversaries will use whatever models/tools they can. They'll target everything with agents 24/7.— Logan Graham, Executive, Anthropic

OpenAI CEO Sam Altman has been similarly blunt. In a conversation with Vanity Fair last month, Altman said the company was willing to accept some things going wrong and warned of a coming tidal wave of cybersecurity problems that are going to come from these open-source models. Joshua Achiam, who resigned as OpenAI's chief futurist this summer after nine years at the company, posted on X: "Feels like March 1, 2020 for Cyber. Not normal, something is happening. Internet lockdowns in a couple of months?"

The anxiety has reached the boardroom. A special Partnership for New York City meeting drew executives from some of the world's largest companies in mid-September, an event that Pfizer CEO Albert Bourla and Tishman Speyer CEO Rob Speyer described as among the organization's biggest gatherings in years. The guests that day were Anthropic whistleblower Jacob Coxon and former Department of Defense official Mark Beall, who addressed an audience of hedge fund managers, bankers, and pharmaceutical executives on the threat posed by AI-enabled cyberattacks. A representative from Coinbase asked what companies could do to combat 'open-weight models in the hands of bad actors.'

Cybersecurity experts caution that the picture is not uniformly apocalyptic — yet. Hamza Chaudhry, the AI and national security lead at the Future of Life Institute, said in an interview with Vanity Fair that AI has so far fallen short of producing a fundamental shift in the scale of harm that attacks can cause. What it has done is lower the barrier to entry for unskilled hackers targeting vulnerable, lower-profile systems. His near-term concern is not a single catastrophic event but a cascade of smaller ones.

Over the coming weeks and months, the more likely scenario is a local hospital network in Ohio losing power and patients on life support dying as a result; a utility network in Texas taken offline; the lights going out across half of Wisconsin.— Hamza Chaudhry, AI and National Security Lead, Future of Life Institute

Chaudhry's deeper worry is what he calls a 'geopolitical Hugging Face moment' — a scenario in which a Mythos-level cyber-capability built by China reaches mercenaries, who turn it on military infrastructure, AI agents veer onto an unintended course, and the chain of events leaves NATO weighing whether to trigger Article 5. His group ran that simulation at the Munich Security Conference earlier this year before an audience of Western national security officials. "There was a real chance that Article 5 would be invoked and NATO would declare war on Russia" in the simulation, he told Vanity Fair.

Adding to the threat is the rapid improvement of freely available, open-weight AI models, most of them produced by Chinese companies. According to Anthropic in September, the strongest of those models now perform at roughly the level Mythos did when it first appeared in April — meaning the capability the U.S. government sought to restrict is effectively in wide circulation. After the PHP vulnerability came to light, Anthropic's deputy chief information security officer, Jason Clinton, offered a stark assessment: "As bad as things have gotten … they are going to get a lot worse before they get better." Eventually, he wrote, AI will help humans write code with no security vulnerabilities — but for now, the industry needs to "hold on."

Why it matters — AI has made sophisticated cyberattacks accessible to amateurs, meaning hospitals, utilities, banks, and government systems everywhere face a threat that no longer requires a nation-state or expert hacker to execute.

Reported by washingtonpost.com, adn.com, vanityfair.com