November 16, 2026 · The AI Honeypots research team


Stolen API keys are only worth something if they still work, and if you know what they work for. Between November 9 and 13, four addresses spent five days finding that out on our decoys. Between them they presented 9,023 distinct API keys in 45,637 requests.

Most keys were tried once. The median key was sent exactly one time, and 8,624 of the 9,023 were only ever sent with one model name. That is not someone using a key. It is someone sorting a list.

The volume shows up fleet-wide. In the week of November 2, our sensors saw 445 distinct credentials that looked like real keys (not placeholders, not our own decoy material). In the week of November 9 they saw 9,169, about 20 times as many. These four addresses account for almost all of them: about 9,000 of the 9,169. None of the 9,023 keys had reached any of our decoys before November.


The four checkers

Address Active (UTC) Distinct keys Peak rate
161.114.15.176 Nov 9 09:38 to Nov 10 02:43 (also active Nov 1 to 7) 3,462 1,882 new keys in one hour (Nov 9, 14:00)
115.60.60.86 Nov 9 14:36 to Nov 11 12:46 3,596 240 requests a minute
165.1.69.195 Nov 10 08:31 to 19:40 1,277 120 requests a minute
115.60.62.106 Nov 12 01:03 to Nov 13 12:38 713 315 requests a minute

Never more than two of them were checking keys at the same time, and each ran for between half a day and two days before going quiet. Sixteen keys turned up at more than one of the four, and two at all four, so at least part of the list was shared. The two 115.60.x.x addresses are in the same network and sent an overlapping set of user agents.

Two of them named their tools. 161.114.15.176 sent four user agents of its own:

Mozilla/5.0 (Windows NT 10.0; Win64; x64) ai-keys-model-liveness/1.0    4,906 requests
Mozilla/5.0 (Windows NT 10.0; Win64; x64) ai-keys-poolfullcat/1.0        1,153
Mozilla/5.0 (Windows NT 10.0; Win64; x64) ai-keys-verify/1.0               546
Mozilla/5.0 (Windows NT 10.0; Win64; x64) ai-keys-verifier/1.0             241

115.60.62.106 sent 2,291 requests as Mozilla/5.0 (compatible; famfill/1.0). We found no public reference to any of these names. "Model liveness" and "pool" describe the job well enough: take a pool of keys and find out which ones are alive, and for which models.


What a check looks like

A check is one short chat request with a cheap prompt and a small token limit. The most common body, sent 3,725 times, was:

{"model": "...", "messages": [{"role": "user", "content": "hi"}], "max_tokens": 24, "temperature": 0, "stream": false}

The prompts are the same kind of test we described last week in How to Spot an AI Honeypot: a one-word answer, a sum, an echo, or a question about who made the model. The most common:

Prompt Meaning Requests
hi 4,117
17+25 等于多少?只回一个数字。 "What is 17+25? Reply with one number only." 3,711
Which company developed you? One sentence. 2,082
说一句话:你属于哪个模型家族? "In one sentence: which model family do you belong to?" 1,939
3+5 等于多少? "What is 3+5?" 1,304
What is 17 plus 25? Reply with the number only. 1,034
Reply with the single word OK. 973

There was also an echo test, 请原样输出这串字符:KQ7XZ9 ("Output this string exactly: KQ7XZ9").

Each prompt answers a different question about the key. "hi" asks whether the key works at all. The sum and the echo ask whether a real model is behind the server or a canned reply. "Which company developed you" and "which model family" ask whether the model is the one the server claims, which matters when the key will be resold as access to a particular model. It is the same check a buyer runs on a cheap API reseller.


Every path, every port

The checkers didn't assume where the API lived. Against the same servers they tried:

  • /v1/chat/completions (32,056 requests) and /chat/completions (7,684)
  • /v1/v1/chat/completions (2,902). That's what you get when a tool appends /v1/chat/completions to a base URL that already ends in /v1. It runs every key against both forms rather than working out which is right.
  • /api/v1/chat/completions, /api/chat/completions, /openai/v1/chat/completions
  • Anthropic-style /v1/messages and /api/messages
  • Gemini-style /v1beta/models/<model>:generateContent
  • GET /v1/models (1,346), to see which model names a server offers before trying keys against them

Keys went to every kind of decoy we run, not just the OpenAI-compatible ones: LiteLLM on port 4000, Open WebUI on 3000, AnythingLLM, KoboldCpp, text-generation-webui, vLLM, Ollama, LM Studio, and even Elasticsearch on port 9200 (1,083 requests). Every service on the target address gets every key.


Asking for models by reseller prefix

Alongside plain model names, the checkers asked for models under routing prefixes that no AI provider uses:

Prefix Requests, Nov 9 to 15 Addresses Distinct model names
ag/ 3,365 34 8
cc/ 1,268 24 8
oc/ 425 3 21
codebuddy/ 309 7 6
ninjasatwork/ 292 11 2
opencode/ 227 3 14
custom_endpoint/ 196 6 6
codex/ 136 7 2
kiro/ 79 7 6

Each prefix was followed by an ordinary model name, mostly Claude models plus some Qwen. Several are the names of coding-assistant products (Codex, OpenCode, CodeBuddy, Kiro). The pattern is consistent with self-hosted proxies that turn coding-assistant subscriptions into an OpenAI-compatible API and route requests by a per-account prefix; at least one such open-source proxy supports exactly that kind of prefix routing. We can't confirm which tool each prefix belongs to. What the requests do show is that key checkers now treat "a subscription behind a proxy" as something worth finding, alongside ordinary provider keys.

This is new. Across the whole fleet, requests with these prefixes went from 1 in the week of August 31 to 895 from 58 addresses in the week of October 26, 823 from 22 addresses the week after, and 6,297 from 51 addresses in the week of November 9. One checker drove most of the jump: 115.60.62.106 sent 4,063 of those requests, under 57 different prefixed names.


If you run an AI gateway

The checks look nothing like a user, which makes them easy to spot in logs:

  • One source, many keys, one request each. A user has one key, or a few. A checker presents hundreds of distinct bearer tokens in an hour, and most of them once. Alert on the count of distinct Authorization values per source address per hour; our four checkers ran between 713 and 3,596 each.
  • Tiny, checkable prompts. max_tokens of 24 or less, temperature 0, and a prompt like "hi", a sum, an echo string or "which company developed you".
  • Path variants against the same host. /v1/v1/chat/completions next to /v1/chat/completions, /chat/completions and /api/v1/....
  • Model names your server doesn't serve. Especially names with a routing prefix like ag/, cc/ or kiro/, if you don't use them.
  • The user agents above. They are cheap to change, but they were in use this week.

If your server accepted one of these requests with a real upstream key behind it, assume that key is now on a list. Rotate it, and check your provider's usage logs for traffic you don't recognise.


Indicators

IP Seen (UTC) Activity
161.114.15.176 Nov 1 to 10 (key checks Nov 9 to 10) 3,462 distinct keys; user agents ai-keys-model-liveness/1.0, ai-keys-poolfullcat/1.0, ai-keys-verify/1.0, ai-keys-verifier/1.0
115.60.60.86 Nov 9 to 11 3,596 distinct keys; up to 240 requests a minute
165.1.69.195 Nov 10 1,277 distinct keys
115.60.62.106 Nov 12 to 13 713 distinct keys; user agent famfill/1.0; 4,063 reseller-prefix model requests

The other user agents these addresses sent were generic: curl/8.0, Python-urllib/3.12, and Chrome- or Windows-style Mozilla/5.0 strings.


The AI Honeypot Observatory runs decoy AI infrastructure and publishes what reaches it. We never test, use or publish the keys we receive. Attacker IPs from every post are in our free feeds at ai-honeypots.com/feeds/.