Independent research, published as it happens.
A small distributed honeypot network, run independently and published without commercial filters. Sensors observe scanners and operators across SSH, HTTP, and AI-infrastructure endpoints. Write-ups are for the defenders working the same surface.
- Events observed
- …
- Unique IPs
- …
- Countries
- …
- IPs reported
- …
A decoy server
that records everything.
A honeypot looks like a real server: an exposed shell, an open API, an unsecured endpoint. Anyone who reaches it is expecting a soft target.
What they have found is an instrument. Every keystroke, every credential, every command is preserved. The operator sees a convincing system. We see what they did.
Across every surface we expose. Around the clock.
Command streams
Brute force, credential stuffing, post-exploitation, persistence, cryptominer deployment: every keystroke timed and preserved.
LLMjacking · MCP
Unauthorized inference on Ollama/OpenAI endpoints. Autonomous agents discovering and invoking MCP tools. Prompt injection. Model enumeration.
HTTP reconnaissance
Web fuzzing, API abuse, Docker registry scanning, credential-file hunting. Full access logs with request, body, and user-agent detail.
Auth patterns
Username/password patterns, SSH key injection, credential family taxonomy, automated-vs-manual targeting signals.
IPs change.
Fingerprints don't.
Every connection produces identifiers derived from how a client communicates. Not what it says.
SSH algorithm negotiation fingerprint. OpenSSH, Paramiko, libssh, and Go's x/crypto each produce distinct values, identifying the tool regardless of source IP.
HTTP client fingerprint based on TLS behavior, header order, and request characteristics. Different libraries produce different values, fingerprinting the scanner itself.
Passive OS fingerprinting from TCP packet characteristics. TCP window sizes, options, and flags reveal the operator's OS without sending a single byte back.
Capture. Classify. Enrich.
Capture
Sensors across four continents, multiple protocols, 8+ services. Every interaction logged with passive fingerprints and microsecond timing.
Classify
Mapped against MITRE ATT&CK. Scored for automation and novelty. Grouped into campaigns. AI agents detected separately.
Enrich
Cross-referenced with AbuseIPDB, GreyNoise, VirusTotal, Shodan. Malicious IPs reported back to the community. Nothing monetized.
Our sensors run a fork of Beelzebub, an open-source low-code honeypot framework by Beelzebub.AI. Every part of our collection pipeline runs on top of their work.
Four continents, constant watch.
Loading sensor health…
Every IP, cross-referenced.
AbuseIPDB
Community-sourced abuse confidence scoring. Every malicious IP we observe is reported back.
GreyNoise
Internet-scan classification. Separates targeted activity from background noise.
VirusTotal
Multi-engine malicious reputation. Cross-references 90+ security vendors.
Shodan
Operator infrastructure fingerprinting: open ports, CVEs, hostnames.
Research integrity,
earned line by line.
- Evidence over speculation.
- Every finding rests on direct session-level evidence. Conclusions stay inside the data.
- Confidence-graded.
- Every judgment is rated high, medium, or low, and the rating is stated.
- Passive collection only.
- No active scanning, no exploit delivery, no rDNS from sensors.
- No false attribution.
- Geography is reported, never blamed. Origin is not attribution.
- Safety first.
- Credentials, working exploits, and infrastructure are all redacted before publication.