# Analysis traces discounted resold Claude access to AWS account farms

> Hans Dekker reverse-engineers an open-source AWS signup tool that he says powers resold Claude access advertised at 6% of Anthropic's API rate.

- Status: reported (One outlet, not independently verified yet.)
- Topic: Safety & Security
- Published: 2026-11-12T09:26:58+00:00
- Updated: 2026-11-12T10:10:29.516705+00:00
- Web page: https://ai-news.ikangai.com/story/analysis-traces-discounted-resold-claude-access-to-aws-account-farms/
- Seen in: Demetrios from the AAIF
- Change: security (Security)
- Applies to: Claude, Kiro, AWS Builder ID, CodeWhisperer endpoint

## What changed

Resold Claude access is backed by free AWS Builder ID accounts that an open-source tool mass-creates through a scripted reimplementation of AWS signup.

Resold Claude access advertised at 0.06× Anthropic's API rate is backed by free AWS Builder ID accounts that an open-source tool mass-creates, Hans Dekker reported on November 5, 2026. Dekker reverse-engineered the tool, an Electron app called Kiro-account-manager published under the GitHub account 'chaogei'. The accounts give access to Claude through Kiro, Amazon's IDE. Sellers in Chinese-language Telegram channels market them at six cents on the dollar. Dekker's central point is that the free tier of an AI product is backed by a frontier model, which gives attackers a strong economic incentive to extract that value. Pooled free accounts can bypass per-account usage limits, be resold through relays, or supply high-volume outputs for distillation. Dekker describes distillation as harvesting a model's answers and reasoning to train a model of their own.

Dekker says the accounts are the same ones buyers ask about. In those channels, buyers ask how to log into a 'kiro' account 'with the Builder ID'. Mentions of kiro rose from a few hundred a month in spring 2026 to tens of thousands a month by midsummer. He describes a layered market. Farmers create accounts, relay operators pool them into shared key groups, and relays sell discounted Claude API access. One seller's listing advertises the pool as supporting enterprise distillation. Some relays sell distillation-only pools with reserved capacity and a minimum spend.

The tool does not drive a browser. Dekker writes that it reimplements AWS's signup as a scripted sequence of more than a dozen HTTP requests. The first request registers an OAuth client under the name of Amazon's command-line tool, Amazon Q Developer for command line, so the traffic looks as if it came from that tool. The one step outside AWS is the email verification code. The tool gets that code either from throwaway addresses on domains the operator owns, which forward to a single inbox on a disposable-mail service, or from real mailboxes the operator supplies, such as bulk Outlook accounts or dot-variants of one Proton address.

The obstacle that follows is Amazon's anti-bot script, called FWCIM. It runs on the sign-in page, collects device and behavior data into a blob, encrypts it with XXTEA and sends it to AWS. Dekker found that the tool fills in the same schema with invented values. It reads the encryption key out of AWS's public app.js at runtime, with a hardcoded fallback. That detail matters for defenders. Because anyone can read that key, the encryption proves only that the blob arrived intact, not that a real browser produced it. Dekker's conclusion is that the fingerprint is untrusted input.

Dekker also describes how the tool hides where the accounts come from. The operator configures a residential proxy, which is the address AWS sees. The supported residential proxies reject mainland-China source IPs, and the tool targets Chinese-language users. So an upstream proxy can be placed in front of the residential one. For use, the tool runs a local reverse proxy that exposes the standard OpenAI and Anthropic APIs. For each request, it picks a farmed account, makes the request look as if it came from the Kiro IDE, and forwards it to AWS's CodeWhisperer endpoint. Any tool built for those APIs works against the relay, so the pooled free accounts behave like a single paid API key.

Dekker finds concrete errors in the forged values. The tool draws the sine, cosine and tangent of a fixed constant from a list of ten options. Only one matches what Chrome's V8 engine returns in his test. Four are arithmetically impossible, since in those rows the squares of sine and cosine sum to 1.0013724361669656 instead of 1. The graphics card string is another case. The tool draws it from a pool of twenty-two hardcoded strings, and twenty-one of them lack the hexadecimal hardware ID that the claimed Windows Chrome would report. The plugin list is shuffled per account, although Chrome has used a fixed list of five PDF entries in a fixed order for years. Behavior is invented too. The email-field click always ends in .5. Two timings are constant on every account: exactly 38 milliseconds on page load and exactly 45 seconds on email verification.

Signals also appear at use time. The user agent string is built from the host the relay runs on. Dekker says a pool on one Linux server reports the same operating system, kernel and Node version for every account, even though each account was created as a Windows machine. The tool assigns each account its own machine ID, but those IDs usually share one exit IP. The result is many machine IDs sending requests from one address with identical environment strings.

Dekker's defender advice rests on a distinction between cheap and costly signals. The inconsistencies are cheap to check, and one caught account can lead to others through patterns in email names and use-time IPs. He argues that a careful operator can fix those specific mistakes, which is why the more durable signal is the sharing. A static residential IP for each account has a fixed cost per address, so the operator runs many accounts through a few shared IPs. The same reasoning applies to email domains and to the batches in which accounts are made. Dekker also warns that if a product is reachable through a partner's client, the partner's checks are the ones being tested.

Dekker's piece does not tie these accounts to any particular buyer or lab. Anthropic's February 23, 2026 report names DeepSeek, Moonshot and MiniMax as running distillation campaigns with over 16 million exchanges through approximately 24,000 fraudulent accounts. Qian's ChinaTalk analysis, published May 5, 2026, covers transfer stations. These are a grey market in which Chinese developers reach Anthropic's models through overseas API proxies, often at as low as 10% of the official price and paid in RMB via WeChat or Alipay. Several Chinese developers, as Qian reports, say the markup is customer acquisition and the harvesting of user logs is the real margin. Qian notes that whether proxy operators systematically sell those logs remains unverified.

## Why it matters

Companies using Claude APIs face abuse risk from resold access, and the listed signals help security teams detect it.

## What you can do

Screen account activity for the forged fingerprints and shared IP addresses listed in the analysis.

## Original source

- [Detecting and preventing distillation attacks \ Anthropic](https://www.anthropic.com/news/detecting-and-preventing-distillation-attacks)
- [How fraudulent AWS accounts feed a Claude distillation market](https://hansdekker.dev/writing/fraudulent-aws-accounts-claude-distillation/?nlb=8biAoiutDp)

## Further reporting

- [How to Buy Cheap Claude Tokens in China - by Zilan Qian](https://www.chinatalk.media/p/how-to-buy-cheap-claude-tokens-in)

---

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it.

Cite as: IKANGAI AI News, "Analysis traces discounted resold Claude access to AWS account farms", 2026-11-12, https://ai-news.ikangai.com/story/analysis-traces-discounted-resold-claude-access-to-aws-account-farms/
