Hackers (with a little help from Claude) launch one-click cyberattacks

Just a heads up, if you buy something through our links, we may get a small share of the sale. It’s one of the ways we keep the lights on here. Click here for more.

Anthropic has found itself in an awkward new plot twist: its own AI, Claude, ended up helping Chinese state-backed hackers pull off a string of cyberattacks, and it did so with startling efficiency. 

According to reporting from The Wall Street Journal, the hackers used Claude to automate roughly 30 attacks on companies and governments during a single campaign in September.

How automated are we talking? Try 80 to 90 percent. 

Jacob Klein, Anthropic’s head of threat intelligence, told the Journal that the operation essentially ran “with the click of a button,” with humans stepping in only for tiny decision points, moments like “Yes, continue” or “Uh, Claude, are you sure about that?”

It’s the sort of setup that sounds less like a spy thriller and more like an overly helpful intern accidentally aiding international espionage.

Unfortunately, AI-assisted hacking isn’t some shocking outlier anymore. It’s quickly becoming the norm. 

Just this month, Google revealed that Russian hackers were using large-language models to spit out malware commands, the digital equivalent of having ChatGPT ghostwrite your cybercrime.

The US has long warned that China is using AI to steal American data, something China denies. But Anthropic told the Journal it’s confident this particular crew was state-sponsored. 

In the September campaign, the attackers successfully stole sensitive data from four victims. 

As usual, identities remain undisclosed, although Anthropic emphasized that the US government was not one of the successful targets.

The episode underscores a growing problem in tech: AI models can supercharge productivity, creativity, and, unfortunately, hacking. 

Stitching together the dozens of tasks involved in cyberattacks once required skilled operators and lots of time. 

Now? A language model can orchestrate the whole workflow like it’s planning a weekend road trip.

Anthropic says it’s continuing to build guardrails to prevent this kind of misuse. 

But as AI gets more capable, one thing is becoming very clear: the future of cybersecurity might involve fighting off attackers who don’t need hackers at all, just a model and a mouse click.