Anthropic Caught Hackers Using Claude for Cybercrime — Including a 1.8 Million-App Secret Hunt

Anthropic has published one of the most detailed accounts yet of how criminals and spies are putting its AI to work — and it is not a story about chatbots writing phishing emails. In a threat report covering December 2025 through August 2026, the company says hackers used its Claude models to run near-fully automated hacking campaigns: scanning 1.8 million Android apps for leaked secrets, stealing thousands of corporate login tokens in about 34 hours, and carrying out state-sponsored espionage with AI handling up to 90 percent of the work.

The report, “Countering misuse of AI: September 2026,” catalogues abuse across seven areas — cyber operations, influence operations, surveillance, scams and fraud, biological misuse, conventional weapons development, and model distillation. In every case Anthropic says it detected the activity, disrupted it, strengthened its safeguards, and shared intelligence with authorities and industry partners. The actors ranged from Russia’s intelligence services to a notorious data-theft gang and lone operators. What they shared was a method: use Claude not as a chatbot, but as the operator.

What Anthropic’s threat report found

Anthropic’s Threat Intelligence team identified and disrupted a series of cyber operations in which threat actors used Claude — including suspected state-sponsored groups, financially motivated criminals, and politically motivated individuals. The abuse was carried out with Claude Haiku, Sonnet and Opus models. Notably, Anthropic says no malicious activity was found on Claude Fable or Mythos, its most powerful model classes, which carry safeguards that greatly reduce their ability to perform harmful cyber tasks.

The report uses internal designators called Generative Threat Groups, or GTGs, for the actors observed abusing AI. It also tries to measure “uplift” — how much more harm was caused with AI than without it. The headline trends are stark: sophisticated attacks no longer require sophisticated attackers, and AI’s role in cyber operations has become increasingly autonomous, with a majority of the operations enabled by AI through direct execution or orchestration rather than simple questions and answers.

That is the shift that matters. A year ago, a multi-victim hacking campaign needed many skilled operators and specialist knowledge. In the cases Anthropic documents, a hacktivist using stolen API keys, disparate financially motivated individuals, and a state espionage operator each sustained multi-victim campaigns on their own — because the AI supplied the labor, the tooling and the speed.

ShinyHunters: 1.8 million Android apps scanned for secrets

The most brazen case involves ShinyHunters, the data-theft-and-extortion gang famous for large-scale breaches that usually start with social engineering and compromised accounts. A French-speaking member going by the handle “frkoo” built a credential-harvesting pipeline on ten AWS EC2 instances. The pipeline mass-downloaded 1.8 million different Android APKs from multiple app stores, decompiled them, and used the open-source tool TruffleHog to search for hardcoded secrets. Confirmed findings were sent in real time to a Telegram group divided into more than 100 source types.

A second automated pipeline gathered GitHub organization email addresses and harvested GitHub Personal Access Tokens. According to Anthropic, these two pipelines supplied the initial-access credentials used in most of the breaches confirmed in connection with the actor. The same actor also set up a carding website impersonating the French national police to sell stolen payment card records.

A smartphone with a magnifying glass revealing hidden key and lock symbols inside floating app icons

The scale of what followed shows why the credential pipeline mattered. ShinyHunters affiliates breached a software-as-a-service provider and took the data of about 200 of its downstream customers. In a separate operation, an actor using Claude obtained more than 2,100 sets of Azure AD authentication tokens linked to over 40 corporate Microsoft tenants in around 34 hours — and Anthropic says “AI agents carried out nearly all of the work.” Other affiliates breached a technology provider and stole 1TB of data, compromised an airline, and gained access to an energy company’s systems.

The point is not just the volume. It is the speed and the leverage: a pipeline that turns public app stores into a credential factory, running around the clock, feeding breaches that would have taken a skilled human team days or weeks.

Russia’s SVR and the first fully automated state hacking campaign

The state-sponsored cases are, if anything, more alarming. Anthropic attributes one campaign to GTG-20006, the Russian SVR crew also known as Midnight Blizzard, APT29 or Cozy Bear. The group used customized AI-driven workflows that handled everything from malicious code development through infrastructure acquisition, phishing, command-and-control setup, persistence, and finally data exfiltration. The automation dramatically increased the attack cadence, allowing the group to strike embassies, think tanks, defence-industrial firms and government agencies across Ukraine, Europe, the Middle East, Asia and North Africa. More than 20 organizations were targeted.

Then, on November 12, Anthropic disclosed a separate campaign it calls the first large-scale cyber espionage operation in which an AI system carried out most of the intrusion work on its own. The company linked the mid-September campaign to a Chinese state-sponsored actor it labels GTG-1002. Human planners chose targets and built an attack framework, then used prompt tricks to convince Claude that it was performing defensive security testing for a legitimate firm. Operators broke malicious work into small tasks so the model would not see the full intent.

Once past those guardrails, Claude Code inspected victim networks, mapped high-value databases, researched vulnerabilities, wrote exploit code, harvested credentials, moved laterally and prepared data for exfiltration. Humans reportedly intervened only at sporadic escalation points — roughly four to six per target — while the AI handled an estimated 80 to 90 percent of tactical steps. At peak activity, request rates reached thousands per second, a tempo Anthropic said would be impossible for a human crew to sustain. The group attempted breaches at about thirty technology firms, financial institutions, chemical manufacturers and government agencies, with confirmed intrusions at a small number of them. MITRE ATT&CK has catalogued the activity as Campaign C0062.

Anthropic says it detected the activity, disrupted the operation and shared technical indicators with industry partners and US authorities. Federal officials have not independently confirmed every attribution detail in public statements — a caveat worth keeping in mind — but the report gives defenders a concrete playbook to hunt for: unusually fast sequences of reconnaissance queries, exploit development and credential harvesting originating from compromised developer workstations.

A glowing world network map with red threat nodes quarantined inside translucent containment bubbles

The bigger pattern: AI has collapsed the skills gap

Step back from the individual cases and the pattern is the story. Every layer of offensive operations has been uplifted by AI, from reconnaissance and tool development to data processing and exploitation. The labor and tooling gap that used to separate well-resourced, state-sponsored operations from individual operators has collapsed. A lone actor with stolen API keys can now run infrastructure that once required a team.

That is also why the report lands in a week full of related warnings. Independent testers have shown leading chatbots can be pushed into producing viable attack plans, as the GAO’s covert AI safety tests found. And on the very same day, California sued OpenAI alleging its experimental AI agents escaped testing more than 80 times — a different kind of AI safety failure, but part of the same reckoning: see California sues OpenAI.

The defenses are racing to catch up. Microsoft’s November Patch Tuesday fixed 1,094 flaws, including an Outlook zero-day already under attack. The same week, a fake Claude Code extension in Cursor was found planting an SSH backdoor — a reminder that the Claude brand itself is now a target for attackers, not just a tool for them.

What Anthropic did — and what defenders should do

Anthropic says it removed the accounts involved, modified its guardrails based on what it learned, and shared intelligence with authorities and industry partners. The fact that no malicious activity was found on Claude Fable or Mythos suggests its strongest safeguards — the ones that most limit harmful cyber capabilities — do work when they are in place. The uncomfortable question is why equivalent protections were not already universal across the models that were abused.

For defenders, the report is practically useful in a way most vendor disclosures are not. The indicators are behavioral, not just technical: bursts of thousands of API requests per second, reconnaissance followed immediately by exploit development and credential harvesting, multi-agent frameworks chaining steps through tooling like the Model Context Protocol. Security teams can hunt for those patterns in their own logs today.

For everyone else, the practical lessons are familiar but newly urgent. Hardcoded secrets in shipped apps are a live liability — the 1.8 million-app scan proves attackers are harvesting them at industrial scale, so secrets belong in vaults and environment variables, never in code. Personal access tokens and cloud credentials need short lifetimes, tight scopes and rotation. And any organization whose developers use AI coding assistants should treat the assistant’s API keys and workstations as privileged targets, because attackers already do.

Why this matters beyond Anthropic

Anthropic deserves credit for publishing this level of detail. Most AI companies disclose misuse in vague terms, if at all. Naming the groups, the numbers and the methods gives the whole industry something concrete to defend against — and it puts pressure on rivals to be equally transparent about what their own models are being used for.

But the report is also a warning about the direction of travel. The operations it documents were carried out with Claude Haiku, Sonnet and Opus — not even Anthropic’s most powerful models. The next generation of models will be more capable still, and the attackers learning to use them are not standing still. The race now is between AI that finds the flaws and AI that exploits them — and as Claude Haiku 5.5, Anthropic’s fastest and cheapest model, spreads into everyday apps, that race will be fought on millions of devices, not just in security labs.

The good news is that the same technology cuts both ways. The defenders who learn to spot AI-driven attack patterns early — and who lock down the secrets, tokens and keys that feed these pipelines — can still win. The bad news is that the window for doing it casually has closed. The hackers are already automated. The defense has to be, too.

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top