AgentDecoy · Research

What happens to a leaked AI API key: used in 3 seconds, tested on every model

On 25 October 2026 we leaked fake OpenAI keys on purpose, in a public GitHub repo, a public Gist and a Hugging Face Space. Each key pointed back to a server we control, so we could see who used it, when, and what they asked for.

Check whether your own site leaks a key, free

Free. Reads only what any visitor can see. Your score is saved so you can share it; nothing sensitive is ever shown publicly.

How fast

Time from going public to first use:

  • Public Gist: 3 seconds
  • Public GitHub repo: 15 seconds
  • Hugging Face Space: 1 minute 47 seconds

What the tools did with it

Within the first hour the Gist key alone was used by tools on several networks and countries. What they did:

  • listed the available models, the first step of almost every tool
  • sent one-token test prompts to gpt-4o, gpt-4o-mini, gpt-3.5-turbo and even claude-fable-5, assuming a leaked key may sit behind a proxy that serves every model
  • checked the billing endpoint, to see whether the account has credit
  • tried the URL format of Alibaba's Qwen service and a v2 models path, to work out which provider the key belongs to
  • sent our fake server the prompt "Reply with exactly this token and nothing else", a test that the endpoint is a real model
  • sent an Ethereum eth_chainId call to the AI endpoint, which is a crypto node hunter trying every key it finds

It does not stop

One harvester took the Gist's fake AI key and its fake MCP server together, then came back about every 68 minutes, from a new cloud address each time, with the same fake phone identity. Blocking one address does not stop a tool like that. Watching for the pattern does.

What to do if you leak a key

Assume it is already in use. Rotate it at the provider, set a spending limit, and remove it from the repo history, not just the latest commit. Keep keys in environment variables or a secret manager, and check repos before every push.

Questions

How fast is a leaked OpenAI key used?

In our experiment, 3 seconds after a public Gist went live and 15 seconds for a public GitHub repo. Tools watch these sites in real time.

Why did thieves try a Claude model on an OpenAI key?

Many leaked keys sit behind proxies that serve several providers through an OpenAI-style address, so key checkers test a list of models from different companies.

How can I find out if my key is being used?

Plant a tripwire: a fake key with its own address in the same place a real one could leak. AgentDecoy tripwires are free and email you within a minute of use.

Sources: The live Leak Clock, Free tripwires, Repo check GitHub Action

Want it done for you?

Every result lists the fixes with starting prices, or I can look at your site with you. Pro re-checks it every week for $12 a month, 7 days free.

See plans Email Kyle