Claude Code Latest Anti-Ban Complete Guide (July 2026)

After getting 6 Claude Code accounts banned by Anthropic, I’d had enough. I decided to eliminate all Chinese elements and go full US setup: US VPS + US phone number + US Apple Pay, registered a new Claude Code account. It’s been running for almost half a year now, surviving the recent waves of bans.

These past few days Anthropic has been at it again — the timeline is full of complaints, another wave of bans sweeping through.

Asa@app_sail·Jun 30Mystery solved?

Recently the timeline has been full of complaints, many users reporting their Claude Code accounts banned. The cause may have been found.

A hot post on Reddit these days suggests it might be related to Anthropic embedding a backdoor in Claude Code version 2.1.91. Original title: “Anthropic embedded spyware in Claude Code — and attempted to hide it from you”
Show more
From the ClaudeAI community on Reddit
From reddit.com
20933149K

At the same time, a Reddit user decompiled Claude Code and uncovered a detection code. I’ve put together the complete strategy — arguably the most comprehensive anti-ban approach available.

Bottom Line First: A Checklist

Let me go through each item one by one.

Step 1: Change Mac and iPhone Timezone to Taipei

Timezone Masquerade

The decompiled detection logic runs three layers of checks on your local machine:

Layer 1: Read timezone. If it’s Asia/Shanghai or Asia/Urumqi, it triggers additional processing logic targeting the China region. Anthropic is really shady about this, so you need to avoid it.

Layer 2: Check proxy URL. It compares the system proxy URL against a list of 147 entries, including domains from Chinese AI companies like Baidu, Alibaba, Ant Group, ByteDance, Moonshot, MiniMax, Stepfun, etc.

Layer 3: AI Lab keyword scan. It scans the hostname for keywords.

The timezone is the first gate. Once hit, you’re in the risk control queue, and no matter how American your IP, phone number, or payment method is, it’s useless. That’s why this is step one: lowest cost (a few clicks), highest reward (cut off the entire signal chain at the source).

How the encoded result is exfiltrated

The sneakiest part is how the result is sent out. Instead of sending an independent request, it encodes the combined output of the three layers into the system prompt using two nearly invisible details:

• The date changes from 2026-06-30 to 2026/06/30
• The English apostrophe ’ is replaced with one of three visually identical but different Unicode code points (U+2019 / U+02BC / U+02B9)

As soon as the server parses it, it knows your full profile: China + proxy + which AI Lab the proxy points to.

Once this signal fires, the account isn’t banned immediately — it enters the risk control queue. That’s why some people use their accounts for months before getting banned, while others lose theirs in days. The score accumulates until it reaches the threshold.

Solution: Change system timezone to Asia/Taipei

Same UTC+8, display time identical to Beijing, calendar and schedule completely normal — but Taipei isn’t on the detection list.

Some might ask about Singapore. Theoretically possible, but I personally avoid it for two reasons:

• Too many Singapore alts. A lot of domestic accounts selling second-hand Claude/OpenAI accounts use Singapore timezone, plus Singapore itself is a hub for Chinese AI professionals. If Anthropic expands the detection list, Asia/Singapore is the obvious next candidate.
• Taipei is within Claude’s official supported region. Claude officially serves Taiwan, China. Asia/Taipei is a legitimate timezone in their service area, making it the least likely to be actively added to detection.

How to change on Mac

• System Settings → General → Date & Time
• Turn off “Set time zone automatically”
• Click “Set…” and manually search for Taipei
• Important: Go to “Privacy & Security → Location Services → System Services”, turn off “Setting Time Zone” — otherwise Apple will revert it to Shanghai based on your Wi-Fi location

Verify in terminal:

$ readlink /etc/localtime
/var/db/timezone/zoneinfo/Asia/Taipei

$ node -e "console.log(Intl.DateTimeFormat().resolvedOptions().timeZone)"
Asia/Taipei

The second Node command is the API Claude Code uses internally to read the timezone. Returning Asia/Taipei means the masquerade is working.

How to change on iPhone

• Settings → General → Date & Time
• Turn off “Set Automatically”
• Set timezone to Taipei
• Equally important: Settings → Privacy & Security → Location Services → System Services, turn off “Setting Time Zone”

The iPhone step isn’t for Claude Code itself (Claude Code is desktop-only currently), but to prevent iCloud preference syncing between iPhone and Mac from pulling the Mac timezone back to Shanghai. Keeping both consistent is the most hassle-free approach.

Step 2: US VPS

After changing the timezone, the next step is to handle the second layer detection: IP egress. If the egress IP is US-based, Claude Code connects to the “this is a US user” profile.

I use Hostinger, https://www.hostinger.com?REFERRALCODE=AppSail, selecting a US data center. Hostinger is an established hosting provider; reverse lookup of the hostname gives a clean hostinger.com, completely unrelated to that 147-entry list of Chinese AI companies. A machine with 4GB RAM is sufficient; 8GB also works. This machine serves multiple purposes:

• Provides a clean US IP egress
• Hosts a self-built Hysteria2 proxy node (easy setup with OpenClaw)
• Also used to deploy OpenClaw

Avoid Tencent Cloud, Alibaba Cloud for VPS if possible

Some people use “Alibaba Cloud International”, “Tencent Cloud Lighthouse Silicon Valley node”, “Volcengine Overseas” — domestic giants’ overseas data centers. The IP shows US location, which seems fine.

Personally, I think it’s still risky. Even if you only use these machines as VPS relays, the combination of “US egress + Alibaba/Tencent domain” allows Claude Code’s server to infer that a Chinese user is bypassing via a domestic big-tech overseas node. The risk score for this profile could be even higher than direct access from mainland China.

Correct selection:

• Prefer hosting providers with independent domains: Hostinger, CloudCone, Vultr, Linode. Reverse lookup shows hostinger.com — independent domain, not on the list
• IP type is secondary: Residential IP theoretically has a lower risk score than datacenter IP, but as long as the domain isn’t in the 147 list and the IP geolocation is US, datacenter IP works fine. For extra stability, you could add a residential IP proxy layer on top of the VPS. In my experience, expensive residential IP isn’t necessary — Hostinger VPS is good enough.
• If you must use a big cloud provider, choose AWS / GCP / Azure US regions. Even though they are datacenter IPs, at least the domain isn’t on the list
• Absolutely avoid: Any provider with alibaba / aliyun / tencent / qcloud / huawei / volces / bytedance in the name, regardless of “International” or “Global” branding

Configuring Proxy (TUN mode + Clean Residuals)

Once the VPS is set up, this step correctly routes your Mac’s traffic and clears any residual proxy configurations from history. This is the most common point of failure in the entire chain — many people have the VPS ready and timezone changed, only to trip here.

First, figure out your situation

“Does using Clash get me banned?” Two typical scenarios:

• Normal user running Clash / V2Ray / Surge in system proxy mode → Claude Code reads the proxy URL as 127.0.0.1:7890. The 147-entry list doesn’t match, but Layer 1 (timezone) has already hit, so the account still enters the risk control queue. Most individual developers who got banned are in this position.
• Relay service user set ANTHROPIC_BASE_URL → Strictly speaking, this isn’t proxy detection, but the server sees a mismatched Host header and knows you’re using a relay. Independent strong signal.

Common point: as long as Claude Code can read the proxy or base URL configuration, there’s something to scan. Two-pronged solution: TUN mode so macOS network layer sees no proxy; manually clean residuals so toolchain has no base URL leakage.

How to configure Clash on Mac

First, clarify the concepts. Proxy clients have two working modes:

• System proxy mode (likely what you’ve been using): The client sets macOS network preferences HTTP/HTTPS proxy to 127.0.0.1:7890. scutil –proxy reveals it immediately — Claude Code sees it at a glance.
• TUN mode (what we want): The client creates a virtual network interface (utun0) at the kernel level, intercepting all system IP traffic. The macOS system proxy configuration remains empty; scutil –proxy shows all zeros. Claude Code has no awareness of any proxy — it thinks it’s connecting directly to the US internet.

We want TUN mode, not system proxy.

After configuration, verify in terminal:

$ scutil --proxy | grep -E "HTTPEnable|HTTPSEnable"
  HTTPEnable : 0
  HTTPSEnable : 0

$ curl -s ipinfo.io | grep -E '"ip"|"country"|"org"'
  "ip": "104.xxx.xxx.xxx",
  "country": "US",
  "org": "..."

System proxy all zeros + egress IP in US — this level is perfect.

Manually clean historical proxy residuals

TUN mode only solves “macOS network layer doesn’t see proxy”. Environment variables and toolchain residual proxy configurations are a separate problem that must be cleaned independently.

Many people, in the past, added export HTTPS_PROXY=http://127.0.0.1:7890 to ~/.zshrc for npm package installation, or some tool wrote it into ~/.npmrc, .gitconfig — long forgotten, but Claude Code reads all of them at startup.

Go through each command step by step:

# 1. System proxy (should be all zeros)
scutil --proxy | grep -E "HTTPEnable|HTTPSEnable|SOCKSEnable"

# 2. Current shell environment variables
env | grep -iE "proxy"

# 3. Shell startup files for hardcoded proxy
grep -iE "proxy" ~/.zshrc ~/.bashrc ~/.profile ~/.zshenv 2>/dev/null

# 4. npm / yarn / pnpm configuration
npm config get proxy
npm config get https-proxy
cat ~/.npmrc 2>/dev/null | grep -i proxy

# 5. Git configuration
git config --global --get http.proxy
git config --global --get https.proxy

# 6. Claude Code's own settings.json
cat ~/.claude/settings.json 2>/dev/null | grep -iE "proxy|base_url|anthropic_"

# 7. All anthropic / claude related environment variables (especially ANTHROPIC_BASE_URL)
env | grep -iE "anthropic|claude"

# 8. Homebrew
brew config | grep -i proxy

Any entry that is not empty, not 0, not null — manually clear it.

Items 6 and 7 are especially important — ANTHROPIC_BASE_URL is the most commonly overlooked one. If you’ve used third-party relay services (aihubmix, openai-sb, various mirrors) or run Claude Code Router, you may have left a base URL pointing to a non-anthropic.com domain. When the server receives the request and sees the Host header, it knows you’re using a relay.

After cleaning, restart your terminal (to activate the new shell session) and re-run item 2 (env | grep -i proxy) to confirm.

Step 3: US Phone Number

Claude registration and risk control both check phone number region. Domestic numbers won’t work; some number-on-demand platforms pass registration but are easily flagged in subsequent risk control. Strongly recommend using a fixed phone number.

I use RedPocket. Among US phone numbers, Ultra PayGo purple card and Tello are not as cost-effective as RedPocket.

The plan includes 1GB roaming data, uses AT&T network, residential IP — works for everything. Claude Code, TikTok — no problem. Receiving SMS is free, and you can also send 100 international SMS and 100 minutes of roaming calls.

Asa@app_sail·Mar 27 Article
US Phone Number RedPocket: Complete Guide for Application, Activation, and Renewal — More Cost-Effective than Ultra PayGo Purple Card and Tello
Giffgaff UK card has recently been delisted from major platforms like Taobao, JD, and Xianyu, with purchase channels significantly narrowed. If you’re looking for an alternative, RedPocket is currently one of the most cost-effective options, surpassing Ultra PayGo Purple Card and Tello.
RedPocket Mobile… Paid partnership
62170819822K

Step 4: US Apple Pay or US PayPal

Subscription payment info for Claude Pro / Max is another risk control dimension. Domestic dual-currency credit cards or virtual cards either can’t be used for payment or are easily flagged. Recommend linking a US card to US Apple Pay.

Previously I used SafePal’s Swiss bank card, linked to US Apple ID’s Apple Pay. SafePal is no longer available for application in China. Consider other US cards — here are a few recommendations:

• Mercury Bank — suitable for those with a US company
• Capital One — entry-level US card, requires ITIN first
• HSBC US Premier US card

These cards can be linked to US #ApplePay and PayPal. Additionally, US #PayPal can also link Hong Kong cards or mainland China Visa/Mastercard credit cards.

Then link PayPal to your US Apple ID — payments flow smoothly.

Detailed steps on iPhone:

• iPhone: Settings → Apple Account → Payment & Shipping
• Add Payment Method
• Select PayPal, tap “Log in to PayPal”
• Choose United States as country on the bottom left (must be US), language on the right can be set to Chinese
• Fill in the form: email, phone number (only US numbers supported), credit card number (Visa, Mastercard, American Express, etc.)
• Fill in billing address. Refer to screenshot: after entering street address, city/state/zip will auto-complete

You’ll receive an SMS verification code — verify and you’re done.

A Final Word

With this full US setup, my account has been running for almost half a year without issues. Fable 5 is available today — anyone who needs it can start tinkering. If you run into any problems, feel free to comment or join the Discord channel:

https://discord.gg/XDNe7un59V to discuss.

The most uncomfortable part isn’t that Anthropic wants to prevent distillation — that’s reasonable.

What’s uncomfortable is the method: not written in terms of service, not in telemetry documentation. They XOR-obfuscated detection code into the client, encoded the result into a Unicode apostrophe you can’t see, and treated normal users like fools.

This is no longer in the realm of “we collect anonymous usage data” — it’s client-side fingerprinting + covert channel exfiltration.

For individual users, what you can do is clean up your local environment so there are no machine-recognizable Chinese signals left in the client.