2.14 Million Banned Accounts and a 3.3% Appeal Success Rate: How Claude's Anti-Fraud and Ban Protection System Works

• Anthropic's annual ban volume: 2.14 million accounts (690,000 in H1 and 1.45 million in H2, +110%)
• Appeal-to-reinstatement conversion rate: 3.3% (1,700 restored out of 52,000 requests)
• Cascading risk on a shared card: an entire organization of up to 110 employees blocked overnight with no refund
• The main factor in account survival: consistency of IP country, time zone, card, and phone + an email older than 10 years

The statistics published in Anthropic's Transparency Hub have dispelled any illusions about "random glitches" causing loss of access to Claude. In a single year, the system purged 2.14 million accounts: 690,000 in the first half of the year, followed by a jump to 1.45 million in the second half (an increase of more than 110%). Meanwhile, relying on support through the `claude.ai/restricted` form is pointless: of the 52,000 appeals filed in the second half of the year, access was actually restored in only 1,700 cases — a statistical margin of 3.3%.

Starting September 4, 2025, the rules got even stricter: in addition to direct geo-blocking of users from unsupported regions, foreign companies in which founders from such jurisdictions hold more than a 50% stake were also banned. Watching colleagues across the industry lose their working profiles one after another, our team took a close look at how the anti-fraud machine actually works.

Why does

Anthropic's anti-fraud system ban for infrastructure and billing details rather than prompt text?

A common myth holds that bans are triggered by "sensitive" questions in chat. In reality, the primary switch is triggered by financial and behavioral scoring algorithms (similar to Stripe Radar), checking against Acceptable Use Policy clauses that prohibit coordinated activity (`coordinate malicious activity across multiple accounts`) and automated registration.

The system cross-references five technical signals. A single anomaly rarely triggers an instant ban, but the intersection of 2-3 red flags launches automatic account deletion without warning:

1. Entry-point jumps and fingerprint desynchronization: if your IP address hops from Frankfurt to Amsterdam within five minutes, sits in an overused pool of cheap hosting providers, or is open simultaneously in two countries, it triggers the regional-restriction-bypass flag. The anti-fraud system cross-checks the IP against the OS system clock, browser time zone, and font list.

2. Shared payment cards and crypto-virtual cards: the most destructive trigger of 2026. When five developers pay for subscriptions with the same corporate or intermediary card, the algorithm links them into a single fraud cluster and wipes out the entire chain.

A telling case is that of a 110-person agritech company (discussed on Hacker News and in GitHub issue `#51583`), whose entire organization — including its corporate domain — was wiped out overnight with no refund of the charged funds. And when paying with an intermediary's card, any refund on a ban goes to the drop's card, meaning you lose the money permanently. 3.

Google account age: a fresh email address without two-factor authentication (2FA) or service history reads as a bot farm. The paradox in practice: a decade-old Russian Gmail account with real YouTube and Drive history, accessed from a careful foreign IP, survives for years, while a brand-new "European" email gets banned within the first week. 4.

Disposable SMS numbers: virtual numbers from public rental pools are already flagged across hundreds of other registrations and instantly drag a new profile down with them. 5. API traffic anomalies: sharp request spikes, attempts at model distillation (`model scraping`), and mass runs of jailbreak prompts.

How do you set up a bulletproof workstation configuration for direct access?

One of our engineers at Alpina Digital, after a series of frustrating bans, built a strict hygiene checklist that keeps an account running without a single failure:

• Using a primary personal Google account with more than 10 years of history and 2FA enabled.
• Full geo-coherence of billing details under the "one country" rule: a personal Spanish bank card + a personal Spanish phone number + a static Madrid IP address.
• Setting the laptop's system clock strictly to the server's time zone (Madrid) and completely disabling browser geolocation settings.
• A dedicated clean browser used only for working with Claude, plus a VS Code extension (without installing Anthropic's native desktop client, which reads OS telemetry more deeply).
• Forcibly closing all tabs and background processes before switching network routes.

What should you actually do to protect your team from cascading bans?

Running a company's business processes through developers' personal accounts tied to a single foreign card is a time bomb. To keep engineers and product teams from wasting time fighting browser fingerprints, we solved this problem architecturally on two levels: for enterprise clients, we deployed a fault-tolerant API gateway in AlpinaGPT, and for individual and team work with a pool of flagship models — without risking the loss of history and subscriptions — we built the [Nomad AI](https://nomadnet.ai) workspace. We also go through ready-made engineering templates for fault-tolerant routing between Claude Opus and backup models with members of [Vibe Club](https://hamidun.com/club).