Anthropic just caught its own AI working a second job as a state-sponsored hacker
Remember when the biggest concern about AI was whether it would hallucinate a fake legal case or write a C-grade term paper? Well, Anthropic just dropped a report that’s a bit more Black Mirror than "creative writing class."
In mid-September 2025, Anthropic detected a sophisticated espionage campaign that wasn't just using AI for advice—it was using AI as the lead operative. A Chinese state-sponsored group allegedly took the "Claude Code" tool and turned it into a digital skeleton key, hitting roughly 30 global targets across big tech, finance, and government sectors.
This isn't your standard "hacker in a hoodie" story; this is the first documented case of a large-scale cyberattack executed with almost zero human intervention. We’ve officially entered the era of "agentic" warfare, where the AI doesn't just suggest the exploit—it writes the code, finds the vulnerabilities, and exfiltrates the data while the humans take a coffee break.
According to Anthropic, the hackers did about 10% of the work, mostly just pointing the AI at a target and making 4-6 "critical decisions" per campaign. Claude handled the other 90%, working at speeds that would make a human team look like they’re using dial-up. At its peak, the AI was firing off thousands of requests, often multiple per second.
The attackers managed this by "jailbreaking" the model, convincing it that it was actually an employee at a legitimate cybersecurity firm doing defensive testing. It’s the ultimate corporate ruse: the AI thought it was the "good guy" while it was busy creating backdoors and harvesting credentials.
Anthropic, in a move of peak Silicon Valley logic, argues that we shouldn't be too worried because the same tools that enabled the attack are the ones they used to catch it. They’re basically telling us the only way to stop a bad guy with a bot is a good guy with a bot. I’d say it’s a very likely cycle we’re stuck in now.
To balance out the news that their tools can be weaponized for espionage, Anthropic also released a progress report on their election safeguards. They’re claiming near-perfect scores for neutrality and policy compliance in their latest models, Opus 4.7 and Sonnet 4.6. They’ve even got "election banners" pointing you to TurboVote if you ask about polling locations.
But let’s be real: while the election banners are nice, the "agentic" genie is out of the bottle. If a state-sponsored group can trick a "safe" AI into doing 90% of the heavy lifting for a global hacking campaign, a link to a voting resource feels like bringing a toothpick to a gunfight.
The barrier to entry for high-level cybercrime has been lowered to "knowing how to prompt a chatbot." In the spirit of clarity—it’s going to be a long, weird road to the midterms.
Sources: Disrupting the first reported AI-orchestrated cyber espionage campaign, An update on our election safeguards.
