AI Honeypot Observatory
ai-honeypots.com TLP:CLEAR

IOC Feed

Threat intelligence derived from decoy AI services.
Decoys pose as Ollama, LiteLLM, OpenAI-compatible, and LM Studio endpoints.
Updated hourly — all data is TLP:CLEAR (unrestricted, CC0 public domain).

Formats JSON · plain-text IP list · STIX 2.1 · MISP native
License CC0 1.0 Universal
ATT&CK TTPs mapped per indicator
Contact feeds@ai-honeypots.com
Full feed · 14-day free trial, no card

The free feed covers the last 7 or 28 days. The Full feed adds 90 days of history, the actor and campaign API, CSV and the full weekly brief with its IOC appendix.

Available Feeds

7-day IOC feed

All indicators from the past 7 days. Includes actor category, MITRE ATT&CK TTPs, behavioral metadata and analyst context. JSON with full taxonomy.

Latest High-Confidence — 24h

Only high-confidence indicators from the past 24 hours. Suitable for real-time ingestion into SIEMs, firewall blocklists, and threat intel platforms. Refreshes every 15 minutes.

Extended Window — 28 days

Full 4-week history. Useful for threat hunting, retrospective analysis, and building blocklists with broad coverage. Same schema as 7-day feed.

STIX 2.1 Bundle — Platform Ingest

OASIS STIX 2.1 structured bundle containing indicator, intrusion-set, and relationship objects. Compatible with OpenCTI, Anomali ThreatStream, Splunk ES, IBM QRadar, and any TAXII-aware platform. Indicators linked to named campaign intrusion-sets. Content-Type: application/stix+json.

MISP Native Feed

Native MISP feed format with manifest.json and hashes.csv for incremental sync. Add https://ai-honeypots.com/feeds/misp as a feed URL in your MISP instance. Two events available: 7-day and 28-day windows. Each IP is an ip-dst attribute tagged with actor category and MITRE ATT&CK technique IDs.

Get a key or start a trial

Full feed: 14-day trial, no card

On 10 November 2026 the free 7-day feed was at its 1,000-IP cap, which cut it off at IPs with 17 or more hits. With a trial key the same week holds about 5,500 IPs with 2 or more hits, and 90 days hold about 29,000. The key adds 90 days of history, no cap, CSV, the full ranked lists in the API and the live capture-level feed without the 60-minute delay. Prompt text, captured credentials and anything that identifies our decoys are never included.

Using a key

Send it as Authorization: Bearer <key>, as the HTTP Basic password, or in an X-AIH-Key header. Never put it in a URL: ?key= is refused with HTTP 400. A keyed response carries an X-AIH-Plan header, and an unknown or revoked key gets HTTP 401. The free feeds above need no key. Free keys are issued by email on request (contact@ai-honeypots.com); today they get the same data as the free feeds.

Use it in your tools

MISP

For now, import the STIX 2.1 bundle (/feeds/stix.json) with MISP's STIX 2.x import. Our MISP-format feed has a format fix in progress: MISP rejects the event IDs in its manifest. Once it's fixed, add it under Sync Actions > Feeds > Add Feed with the URL https://ai-honeypots.com/feeds/misp, with no trailing slash.

OpenCTI

Import the STIX 2.1 bundle (/feeds/stix.json) with pycti. With a trial key it covers 90 days (the top 200 indicators). Don't point the MISP Feed connector at our MISP feed until its format fix is live. This is client configuration: we haven't run it on an OpenCTI instance ourselves.

Splunk Enterprise Security

Add https://ai-honeypots.com/feeds/iocs.txt?days=28 as a line-oriented threat intelligence download, with the ignoring regular expression (^#|^\s*$), the extracting regular expression ^(\S+)\s+#\s+(.*)$ and the fields ip:$1,description:$2. This is configuration: we haven't run it on Splunk ES ourselves.

Actor Taxonomy

SCANNER-MASS
High-volume single-purpose endpoint scanners; no prompt interaction
SCANNER-ENUM
Multi-persona enumerators systematically mapping AI service surfaces Model-enumeration recon explained
SCANNER-COORDINATED
Multi-IP clusters operating as a single coordinated scan campaign
MCP-SCANNER
Repeated Model Context Protocol requests (3+ and at least 10% of the IP’s activity) or MCP tool calls. Verified research-scanner traffic (Censys, Infrawatch) is excluded. Occasional MCP requests are tagged mcp_probe instead
RESEARCH-SCANNER
Verified internet-census scanner (Censys, by its published and ASN-announced networks; Infrawatch, by networks registered to it in the RIPE database and forward-confirmed reverse DNS). Kept in the feed and labelled via scanner_label; filter it out if you only want threats
RELAY-OPERATOR
Shadow API backend pool managers; clockwork rotation, no prompts How relay operators work
RELAY-VERIFIER
Automated relay health-check systems using canned test prompts How relay operators work
RELAY-CUSTOMER
End-users of relay pools; requests using relay-specific model aliases How relay operators work
RELAY-CATALOGER
Model×persona matrix sweepers building backend capability inventories Model-enumeration recon explained
IDENTITY-PROBER
Prompt-engineering attacks to extract true model identity / bypass system prompts Model-enumeration recon explained
CREDENTIAL-HARVESTER
Targets .env, API keys, config files — no AI model interaction Hardening guide
INFRA-COLLISION
Legitimate infra accidentally hitting AI endpoints (NTLM, Grafana, etc.)

Integration Examples

# Download plain-text IP blocklist curl -s https://ai-honeypots.com/feeds/iocs.txt | grep -v '^#' | awk '{print $1}' > ai_iocs.txt # Python — ingest full feed import requests feed = requests.get("https://ai-honeypots.com/feeds/iocs.json").json() for ioc in feed["indicators"]: print(ioc["value"], ioc["actor_category"], ioc["confidence"]) # jq — filter by actor category curl -s https://ai-honeypots.com/feeds/iocs.json \ | jq '.indicators[] | select(.actor_category=="MCP-SCANNER") | .value' # jq — leave out verified research scanners (Censys, Infrawatch) curl -s https://ai-honeypots.com/feeds/iocs.json \ | jq '.indicators[] | select(.scanner_label == null) | .value' # High-confidence only (last 24h, for real-time blocking) curl -s https://ai-honeypots.com/feeds/latest.json | jq '.indicators[].value' # STIX 2.1 — download and count object types curl -s https://ai-honeypots.com/feeds/stix.json \ | jq '.objects | group_by(.type) | map({(.[0].type): length}) | add' # MISP — add as a native feed (recommended) # In MISP: Sync Actions → Feeds → Add Feed # URL: https://ai-honeypots.com/feeds/misp (no trailing slash) # Format: MISP # Enabled: yes # MISP will auto-fetch manifest.json, hashes.csv, and event files. # MISP — import STIX 2.1 bundle via PyMISP (alternative) from pymisp import PyMISP from stix2 import parse bundle = parse(open("stix.json").read(), allow_custom=True) # then push to MISP via misp.add_event() or use MISP's built-in STIX import # OpenCTI — push via pycti from pycti import OpenCTIApiClient api = OpenCTIApiClient("https://your-opencti/", "your-token") api.stix2.import_bundle_from_file("stix.json")

Feed Schema

{ "ioc_type": "ip", "value": "1.2.3.4", "tlp": "CLEAR", "confidence": "high", // low | medium | high | very-high "actor_category": "MCP-SCANNER", // see taxonomy above "ttps": ["T1595.002","T1190"],// MITRE ATT&CK v20 technique IDs (see attack_mapping) "first_seen": "2024-06-01T00:00:00+00:00", "last_seen": "2024-06-07T23:59:59+00:00", "total_hits": 1842, "distinct_personas": 6, "distinct_paths": 23, "prompt_count": 0, "user_agents": ["python-httpx/0.27.0"], "models_requested": [], "interesting_paths": ["/mcp","/mcp/v1/tools","/mcp/list"], "sample_prompts": [], // always empty: prompt text is withheld (paid tier) "prompt_categories": {"identity-probe": 2}, // derived behaviour labels, counts of distinct prompts "prompt_length_bucket": "20-99", // median prompt length bucket "prompt_text_access": "withheld-public; ...", "details": "Analyst narrative with behavioral context...", "scanner_label": null, // or {"class":"research-census","org":"Censys" or "Infrawatch","basis":"..."} "mcp_probe": true, // requested an MCP path in the window "mcp_tool_calls": 12, // MCP JSON-RPC tools/call requests "source": "ai-honeypots.com", "honeypot_context": "Distributed LLM honeypot — TLP:CLEAR, no PII" }