AI Honeypots

Full feed trial: quickstart

AI Honeypots research team. For Full feed (Pro) trial keys. Every endpoint, header and parameter below was checked against the live service on 10 November 2026.

What the trial adds

Our decoys pose as self-hosted AI infrastructure: Ollama, vLLM, LiteLLM, OpenAI-compatible APIs, vector databases and MCP servers. Every IP in the feed sent requests to them. The public feed is cut down so it can stay free. Your key removes the cuts.

Public feedWith your trial key
History (days)1 to 28 days1 to 90 days
IPs per pullthe top 1,000 by hitsno cap
Minimum hits per IP (min_hits)fixed at 52 to 100
CSV (/feeds/iocs.csv)not availableavailable
Ranked lists in the API (threat actors, scanners, paths, user agents, models, campaign IPs, new actors, 0-Day Radar)top 5the full list
0-Day Radar windowup to 28 daysup to 90 days
Capture-level feed (/api/v1/feed)at least 60 minutes old; no path, user agent or model; IPs masked after the newest 20live, every field, IPs unmasked

On 10 November 2026 the public 7-day feed was at its 1,000-IP cap, which cut it off at IPs with 17 or more hits. The same 7 days with your key and min_hits=2 hold about 5,500 IPs. 90 days at min_hits=2 hold about 29,000.

No plan includes prompt text, captured credential values or anything that identifies our decoys. Prompt and credential fields appear as labels only, for example a prompt category and a length bucket.

1. Store your key

2. Send the key with each request

Use any one of these three. They're equivalent.

MethodExample
Authorization: Bearer header (recommended)curl -H "Authorization: Bearer $AIH_KEY" ...
HTTP Basic, with the key as the password and any user namecurl -u "feed:$AIH_KEY" ...
X-AIH-Key headercurl -H "X-AIH-Key: $AIH_KEY" ...

Check that the key is being sent. A keyed response carries X-AIH-Plan: pro_trial. A response without that header is the public feed.

curl -s -o /dev/null -D - -H "Authorization: Bearer $AIH_KEY" https://ai-honeypots.com/feeds/iocs.json | grep -i '^x-aih'

Keyed responses also carry X-AIH-Data-Delay-Hours (0 on the trial), X-RateLimit-Limit (requests per second) and X-AIH-Quota-Remaining (requests left today, UTC).

3. Endpoints and formats

Base URL: https://ai-honeypots.com

WhatPathFormatParameters with your key
IOC feed/feeds/iocs.jsonJSON: feed metadata plus an indicators arraydays 1 to 90 (default 7), min_hits 2 to 100 (default 5)
IP blocklist/feeds/iocs.txtPlain text: one IP per line, # comment lines at the topthe same
CSV/feeds/iocs.csvCSV with a header row (key required)the same
STIX 2.1 bundle/feeds/stix.jsonapplication/stix+jsonthe same; the bundle carries the top 200 indicators by hits
MISP feed/feeds/mispmanifest.json, hashes.csv and two events (7-day and 28-day)none: the same for every plan. Format fix in progress: MISP can't fetch the events yet, so use STIX for now (section 7)
Last 24 hours, high confidence/feeds/latest.jsonJSONnone: the same for every plan
Capture-level feed/api/v1/feedJSON array, newest firstlimit 1 to 200 (default 50), days 0 to 28 (0 means no window)

How the parameters behave:

JSON indicator fields include value (the attacker IP), actor_category, confidence, ttps (MITRE ATT&CK technique IDs), first_seen, last_seen, total_hits, distinct_personas, distinct_paths, prompt_count, user_agents (up to 5), models_requested (up to 10), interesting_paths, prompt_categories, prompt_length_bucket, scanner_label, mcp_probe, mcp_tool_calls and details. The feed's own taxonomy object defines every category and confidence level.

CSV columns: value, actor_category, confidence, total_hits, first_seen, last_seen, distinct_personas, distinct_paths, prompt_count, ttps (separated by ;), scanner_label (the organisation, when the IP is a verified research scanner), mcp_probe and mcp_tool_calls.

The capture-level feed (/api/v1/feed) gives one row per request our decoys received: time, attacker IP, method, path, user agent, the model asked for, the endpoint family, and prompt and credential labels.

Ranked lists with your key return in full: /api/v1/threat-actors, /api/v1/scanners, /api/v1/toppaths, /api/v1/useragents, /api/v1/models, /api/v1/campaigns (top IPs per campaign), /api/v1/new-actors, /api/v1/new-uris and /api/v1/zeroday-radar. They keep their own limit and days parameters where they have them, with days capped at 28. The 0-Day Radar takes days up to 90.

4. Rate limits

5. Examples

Quote each URL, because & means something to your shell.

# JSON: 90 days, IPs with 2 or more hits
curl -s -H "Authorization: Bearer $AIH_KEY" "https://ai-honeypots.com/feeds/iocs.json?days=90&min_hits=2" -o iocs-90d.json

# CSV: 90 days, for a spreadsheet or a SIEM lookup
curl -s -H "Authorization: Bearer $AIH_KEY" "https://ai-honeypots.com/feeds/iocs.csv?days=90&min_hits=2" -o iocs-90d.csv

# TXT: a 28-day blocklist, IPs only
curl -s -H "Authorization: Bearer $AIH_KEY" "https://ai-honeypots.com/feeds/iocs.txt?days=28&min_hits=2" \
  | grep -v '^#' | awk '{print $1}' > blocklist.txt

# STIX 2.1: 90 days (top 200 indicators)
curl -s -H "Authorization: Bearer $AIH_KEY" "https://ai-honeypots.com/feeds/stix.json?days=90" -o stix-90d.json

# The live capture-level feed: the newest 200 requests
curl -s -H "Authorization: Bearer $AIH_KEY" "https://ai-honeypots.com/api/v1/feed?limit=200" \
  | jq '.[] | {timestamp, source_ip, method, path, endpoint_family}'

# The same JSON pull with HTTP Basic instead of a Bearer header
curl -s -u "feed:$AIH_KEY" "https://ai-honeypots.com/feeds/iocs.json?days=90&min_hits=2" -o iocs-90d.json

6. Three things to try first

These use fields that only decoy telemetry like ours can fill. Each runs on the 90-day JSON file from section 5.

  1. IPs that asked for 10 or more different model names. That is model enumeration: mapping what a server will run. On 10 November 2026, about 840 IPs in the 90-day window did this.
    jq -r '.indicators[] | select((.models_requested|length) >= 10) | [.value, .actor_category, .total_hits] | @tsv' iocs-90d.json
  2. IPs that called MCP tools, not just discovered an MCP endpoint (mcp_tool_calls counts JSON-RPC tools/call requests). There were 9 such IPs in the 90 days to 10 November 2026.
    jq -r '.indicators[] | select(.mcp_tool_calls > 0) | [.value, .mcp_tool_calls, .actor_category] | @tsv' iocs-90d.json
  3. Leave out verified research scanners (Censys and Infrawatch) before you block or alert on anything. They stay in the feed with a scanner_label, so you can choose.
    jq -r '.indicators[] | select(.scanner_label == null) | .value' iocs-90d.json

The four RELAY-* categories (operator, verifier, customer and cataloger) describe "shadow API" resellers that use exposed AI servers as backends. The feed's taxonomy object explains each one.

7. Set it up in your tools

MISP

For every IP rather than the top 200, use the JSON, CSV or TXT feed with your key.

OpenCTI

Import the keyed STIX bundle (90 days, top 200 indicators) with pycti on a schedule. Don't point OpenCTI's MISP Feed connector at our MISP feed until its format fix is live (see MISP above).

import os
import requests
from pycti import OpenCTIApiClient

bundle = requests.get("https://ai-honeypots.com/feeds/stix.json?days=90",
                      headers={"Authorization": "Bearer " + os.environ["AIH_KEY"]}, timeout=300).text
api = OpenCTIApiClient("https://your-opencti.example", os.environ["OPENCTI_TOKEN"])
api.stix2.import_bundle_from_json(bundle)

For every IP rather than the top 200, map the JSON or CSV feed in your own connector: value becomes an IPv4 observable, and actor_category and ttps become labels.

This is client configuration. We haven't run it on an OpenCTI instance ourselves.

Splunk Enterprise Security

Add the TXT feed as a line-oriented threat intelligence download. Menu names differ between ES versions. In recent versions the downloads are under Configure > Data Enrichment > Threat Intelligence Management.

With your key: set the download's remote-site user to any name (for example feed), store the key as that user's password in ES credential management, and change the URL to https://ai-honeypots.com/feeds/iocs.txt?days=90&min_hits=2. Then open one downloaded file and check that its # Window: line says 90 days. Our server never asks for credentials, because anonymous pulls are allowed. So if your ES version only sends a password when asked, the line says 28 days and you're getting the public feed. In that case, pull the CSV on a schedule with curl (section 5) and load it as a lookup or a threat intelligence upload.

This is configuration. We haven't run it on a Splunk ES instance ourselves.

8. When the trial ends

9. Terms and help