Anthropic report says AI agents could make more companies worth hacking

Anthropic’s latest Threat Intelligence report paints a blunt picture: AI is lowering the skill bar for cybercrime, and that changes who gets picked on. When attacks used to take real expertise, hackers went after the richest prizes. Now, the company says, the effort needed to probe a target, find weak spots and write exploit code is low enough that smaller or less famous companies start looking attractive too.

The report says attackers have been using Claude to move through corporate systems they barely understood, hunt for valuable data and generate code to exploit vulnerabilities. In one case, a stolen developer token was enough to give an attacker full administrative control of a company’s cloud setup in about three hours. In another intrusion, attackers Anthropic described as suspected ShinyHunters affiliates took data from roughly 200 customers of a software provider. Anthropic says AI agents did almost all of the work.

The mechanics are ugly and efficient. In that software-provider breach, attackers pulled more than 2,100 Azure AD authentication tokens from more than 40 corporate cloud environments in about 34 hours. Those tokens can let someone into cloud services as a legitimate user, no password required. The hackers gave Claude broad goals, let it write and run scripts, and kept pushing it to try something else whenever a tactic failed. Anthropic calls that “vibe hacking,” and says it has spread from a suspected state-backed campaign it saw last November to nearly every kind of cyber attacker it studied.

The report also describes attackers using AI to coordinate parallel agents, test malware against security tools, and keep rewriting code until it slipped past detection. One Russian-speaking actor with a history of hitting hotel-booking and financial-technology platforms allegedly stole about 26 GB from a victim and sought between $1.5 million and $2.5 million through extortion or dark-web sales. A later campaign from the same infrastructure targeted roughly 30 AI companies in about four days.

And the scam side is just as industrial. Anthropic says a China-based app studio used Claude to run more than 4,700 dating-app personas and talk to at least 25,000 people in two weeks in April. Real workers stepped in for live video calls and social follows to make the operation look real. The report’s larger warning is simple: if AI makes every scam cheaper, hackers and fraudsters stop being picky.