🤚 The Open-Palm Credential Banquet

Anthropic says attackers tried to turn Claude into an operations desk for cybercrime and espionage, because apparently the threat landscape looked at automation and said, finally, concierge service. As reported by BleepingComputer, the company described multiple abuse cases spanning financially motivated crime, Russian-linked espionage, and a Chinese-speaking threat group.

The most cinematic case involved an alleged French-speaking ShinyHunters-linked actor using the handle “frkoo”. Anthropic says the actor distributed a credential-harvesting pipeline across 10 AWS EC2 workers, mass-downloaded 1.8 million Android APKs from multiple app-store sources, decompiled them, and scanned them with TruffleHog for hardcoded secrets.

Verified findings were reportedly routed in real time to a Telegram group organized into more than 100 source types. This is the modern software supply chain in miniature: millions of mobile apps, countless embedded secrets, and one chat room where all the consequences arrive wearing a delivery uniform.

Anthropic says the same actor used another automated process to collect GitHub organization email addresses and obtain GitHub Personal Access Tokens. Those pipelines allegedly supplied initial-access credentials used for the bulk of confirmed breaches associated with the actor. In one case, a suspected ShinyHunters threat actor allegedly used Claude-assisted workflows to extract authentication data and obtain more than 2,100 Azure AD authentication token sets tied to over 40 corporate Microsoft tenants in about 34 hours.

👐 The Two-Handed Reality Check

This is not the cartoon version of AI cybercrime where a chatbot says “hack the planet” and lightning comes out. It is more practical, more boring, and therefore more dangerous. The model is being used to glue together tedious offensive workflows: parsing outputs, writing scripts, organizing evidence, improving prompts, automating recon, and converting scattered access into operational momentum.

That is why the 34-hour figure matters. The scandal is not that AI invented credential theft. Humans managed that perfectly well while wearing hoodies and ignoring patch notes. The scandal is that agentic tooling can compress the labor, sequencing, and iteration required to industrialize the theft. It turns cybercrime from a craft workshop into a politely staffed processing center.

Anthropic also described activity attributed to Midnight Blizzard, the Russian espionage group, involving Claude-assisted automation across malware development, infrastructure acquisition, phishing, persistence, command-and-control, and data exfiltration. BleepingComputer reports that the campaigns targeted more than 20 government, defense, diplomatic, intelligence, and foreign-policy entities.

Another operation, attributed by Anthropic to a Chinese-speaking group tracked as GTG-10007, allegedly used Claude as an engineering and orchestration layer for offensive work. The group reportedly ran autonomous vulnerability-research workflows that discovered previously unknown vulnerabilities in a major security product and produced working exploits for several families of network and security appliances.

🌿 The Gentle Awakening

The uncomfortable lesson is that AI misuse does not need science-fiction villainy. It needs scale. Give attackers a system that can reason through repetitive tasks, generate code, summarize logs, and maintain momentum while the human operator edits the playbook, and you have not replaced the criminal. You have promoted the criminal to manager.

This is especially grim for defenders because the weak points are familiar: hardcoded secrets in mobile apps, over-permissive tokens, exposed developer workflows, compromised cloud accounts, and organizations still treating credential hygiene as a quarterly aspiration. AI did not create these weaknesses. It merely arrived with a silver tray and said, would sir like them harvested faster?

Anthropic says it disrupted the malicious activity, banned accounts, adjusted guardrails, improved detection, and contacted authorities, industry partners, and victims. That is the correct response, and it is also the beginning of a much longer governance problem. Every frontier model provider is now, whether it likes the job title or not, part platform company, part abuse desk, part cyber-intelligence shop, and part parent in a room full of toddlers with soldering irons.

👑 The Gold-Leaf Reckoning

The premium verdict is that defenders should stop asking whether attackers will use AI. That question has now been escorted from the premises for wearing last season’s assumptions. The better question is which parts of the attack chain become fast enough, cheap enough, and reliable enough to change the risk model.

For software teams, the answer starts with the unglamorous aristocracy: remove hardcoded secrets, rotate exposed tokens, restrict personal access tokens, monitor impossible cloud activity, test incident response against machine-speed workflows, and assume that public code, mobile binaries, and forgotten build artifacts are being reviewed by tireless systems with no social life and excellent patience.

For AI companies, the lesson is equally ornamental and brutal: safety policy must include abuse operations at industrial scale. The model is no longer just answering questions. In the hands of adversaries, it can become the clipboard that keeps the heist on schedule.

“The future of cybercrime is not a genius in a basement. It is a mediocre process with excellent automation and a premium API key.” — The Slap of Wisdom Threat Intelligence Sommelier Desk, pairing stolen tokens with notes of institutional fatigue