Anthropic: hackers used Claude to steal secrets from 1.8M Android apps

Anthropic has published a report on attempts to abuse its Claude model for criminal purposes. Between December 2025 and August 2026, its specialists documented various forms of such use: cyber operations and influence operations, surveillance, fraud, as well as weapons development and model “distillation.” Some of this activity was disrupted, and the associated accounts were blocked.

Mass harvesting of secrets

One of the most telling cases involves the ShinyHunters collective, known for major data thefts. A suspected French-speaking member of the group, going by the handle frkoo, deployed a credential-harvesting pipeline across ten AWS EC2 workers. The pipeline downloaded 1.8 million individual Android apps from several stores, decompiled them, and used the TruffleHog tool to scour the code for hardcoded secrets. Confirmed finds were sent in real time to a Telegram group split into more than a hundred source categories.

The same person collected corporate addresses from GitHub and obtained personal access tokens (PATs). It was these two pipelines that accounted for the “majority of confirmed breaches.” In addition, frkoo opened a carding shop on a domain masquerading as the French national police, and sold stolen payment card data right down to an interactive map of victims' addresses.

Attacks at the speed of AI

The pace is especially alarming. According to Anthropic, it took one ShinyHunters member roughly 34 hours to extract more than 2,100 sets of Azure AD authentication tokens from over 40 corporate Microsoft tenants — with “almost all of the work performed by AI agents.” In another episode, an attacker went from a single stolen developer token to full administrative control in under three hours.

Russia and China

The report singles out the Russian espionage group Midnight Blizzard, which used Claude to automate malware development, phishing, persistence in systems, C2 management, and data exfiltration. Notably, the group set up a “feedback loop”: as soon as security products detected the malware, it was rebuilt from scratch. More than 20 government, defense, and diplomatic entities were targeted, and the operators mostly just tweaked ready-made Claude Code scripts.

The Chinese-speaking group GTG-10007 used Claude as an “engineering and orchestration layer” for an offensive program: reconnaissance of networks in the Middle East, Europe, and Southeast Asia, and the development of exploits against popular endpoint security solutions. Autonomous vulnerability-hunting processes ran while the operators were away and uncovered several previously unknown flaws. In all, around 50 organizations were hit.

Anthropic says it has strengthened its safeguards, added tools for early detection of abuse, and notified the authorities, industry partners, and those affected.

Source: Bleeping Computer