CyberHalluciNet

Fail-closed decoy sensor

A.I. Decoy Engine

The CyberHalluciNet Decoy Engine is an attacker-facing sensor fabric: multi-protocol surfaces that use constrained AI to stay believable, engage adversaries longer, and feed CTI, while fail-closed policy keeps models and attackers from gaining real authority.

Interactive tutorials Protocols Install

A.I features

What the A.I Honeypot can do with a constrained broker: and where to learn each skill.

Responses

Answer with AI

Emulated shell/HTTP/VFS slots (shell.stdout, http.index_body, …). Policy still validates every byte.

Start: Make the A.I Honeypot

Data fills

Fill decoy data with AI

Redis, SQL, Mongo, Elasticsearch bait via schema-bound seeds (*-ai-seed=1) on Emulated Listen surfaces.

Tutorial: DB AI bait

Camera

Hikvision photos with AI

Optional Ollama scene JPEG pool for App-webs / ISAPI snapshots (CYBERHALLUCINET_HTTP_HIKVISION_AI_IMAGE=1).

Tutorial: Hikvision Path B
  • Schema-bound AI content Provider-neutral broker suggests fills; Sensor policy retains authority. Model output cannot choose OS execution, egress, or tier transitions.
  • How to turn it on Emulated tier + MODE=ai + broker UDS, then choose an adapter (static_stub → ollama / hosted). Full path: A.I Honeypot tutorials.
  • Agentic attack awareness Observe / Probe detection for LLM-assisted attackers, separate from content fills. Scores never auto-block production networks.
  • Safe degradation Broker or AI unavailable → static personas / templates or shutdown, never expanded access.

Interaction tiers

Operator-selected, fail-closed, strictly additive.

Contained · default

Low interaction

Multi-service synthetic decoys. No real shell. No attacker egress. Fastest, safest default.

Emulated

Medium · A.I depth

Rich synthetic FS + schema-bound AI fills and replay. Still no real OS command authority.

Live · ack-gated

High interaction

Sandboxed proxy to disposable guests (MicroVM / pool). Real execution only after signed ack.

Supported protocols

Full sensor listen surface (opt-in, empty = off, fail-closed by default). Yes: mail, DHCP, Windows, cloud, and OT faces are included.

IT & remote access

  • SSH
  • Telnet
  • HTTP/1.1
  • HTTPS
  • VNC / RFB
  • MCP decoy (canary)

Mail & name services

  • SMTP
  • POP3
  • DNS

File, print & Windows

  • FTP
  • TFTP
  • SMB / CIFS
  • MS-RPC (EPM)
  • IPP (print)

Network & voice

  • DHCP
  • SNMP
  • SIP (UDP + TCP)

Databases

  • Redis
  • Elasticsearch
  • MongoDB
  • MySQL
  • MSSQL / TDS
  • PostgreSQL

IP camera / IoT

  • Hikvision App-webs / ISAPI
  • RTSP
  • Hikvision SDK
  • Optional AI / scene images

Cloud-native & AI supply chain

  • Cloud metadata / IMDS
  • Kubernetes API
  • Docker Engine API
  • OCI / Docker Registry V2
  • OAuth device-code lure
  • Model registry lure

OT / ICS

  • Modbus TCP
  • EtherNet/IP
  • Rockwell CSP / DF1
  • S7comm
  • BACnet/IP
  • IPMI / RMCP
  • DNP3
  • IEC 60870-5-104
  • Cellular gateway HTTP
  • MicroLogix / water-WWS pack

Listeners are enabled per address flag / YAML (for example --listen-smtp, --listen-pop, --listen-dhcp). Additional persona packages and lab faces can extend realism; OT listeners share a unified device persona for coherent CTI correlation.

Options

Choose AI posture, placement, and industry pack.

AI detection modes

  • Off Default fail-closed, no agent feature extraction.
  • Observe Advisory telemetry only, never changes attacker-visible responses.
  • Probe Inert markers / lures for secondary confirmation (config-driven, not score-driven).

Placement

  • Research loopback Local Contained lab on loopback / Compose: evaluation and purple-team drills.
  • Internet / CTI decoy posture Fleet max-tier policy and sticky per-source admission keep Internet-facing exposure under operator control.
  • Enterprise internal planting Ops Launch / Investigate / Manage for internal-network intelligence and authorized response.
  • Industry packs Optional packs (water/wastewater OT, healthcare, retail, logistics) for faster persona realism.

Easy to install

From package to Ops UI in about ten minutes. Compose-only path if you prefer containers first.

Path A: Ops GUI (recommended)

  1. Obtain the CyberHalluciNet lab package
  2. make ops-web build
  3. Run ./bin/ops-gui -doctor
  4. Start Ops on 127.0.0.1:3080
  5. Setup → Launch → Investigate / Manage

Path B: Compose research lab

  1. make research-up
  2. make research-self-test
  3. make research-down when finished

A Helm chart scaffold for sensor deploy is included for Kubernetes trials. Live / MicroVM paths need extra ack and host prerequisites.

Quick start commands

# From your CyberHalluciNet lab package root:
make ops-web build
./bin/ops-gui -doctor
CYBERHALLUCINET_REPO_ROOT="$PWD" ./bin/ops-gui -addr 127.0.0.1:3080
# open http://127.0.0.1:3080/ → Setup (Intel + Detection) → Launch → Investigate / Manage

Docs: Install lab · interactive tutorials · Back to platform overview

Learn by doing

Interactive tutorials

Pick a learning path. Steps are ordered the same way as the in-repo Diátaxis tutorials. Progress is saved in this browser; open each tutorial here when you are ready.

Answers

Frequently asked questions

What is the CyberHalluciNet A.I. Decoy Engine?

The A.I. Decoy Engine is CyberHalluciNet’s attacker-facing sensor fabric. It exposes multi-protocol decoy surfaces that use constrained AI to stay believable, engage adversaries longer, and feed CTI, while fail-closed policy keeps models and attackers from gaining real authority over execution or egress.

Which interaction tiers does the Decoy Engine support?

Three tiers: Contained (low-interaction synthetic Listen), Emulated (medium-interaction with schema-bound AI fills and richer personas), and Live (high-interaction, acknowledgment-gated proxy into disposable guest backends). Contained and Emulated never execute real attacker OS commands on the host.

How do I get started with the Decoy Engine safely?

Start on loopback with the beginner lab or Make the A.I. Decoy Engine tutorial. Use Docker Compose or Helm for isolated staging. Do not Internet-bind research profiles by default. Live / MicroVM paths require signed acknowledgment and burnable, network-isolated guests before any higher-risk engagement.

Safety model

  • No attacker authority Attacker input and model output cannot select execution or egress.
  • Broker-only inference Only the policy engine may call the AI broker.
  • Telemetry ≠ policy Detection scores never mutate firewall/IAM without human approval.
  • Research ≠ production Loopback research profile is not an Internet-safe claim by itself.