What this is

GHOSTPROTOCOL is a deliberately tiny chat model packaged for red team and adversarial research workflows where a full-size LLM is the wrong tool — offline boxes, air-gapped labs, implant-adjacent tooling, CI harnesses, throwaway containers, and anywhere you need a local model that loads in under a second and costs nothing to run.

It is 91 MB. It runs on a laptop CPU, a Raspberry Pi, or a dropbox. It does not need a GPU and it never leaves the host.

Callsign: GHOST.

Quickstart

ollama run zerodaylk/GHOSTPROTOCOL

>>> Who are you?
I am GHOSTPROTOCOL. Callsign GHOST.

API

curl http://localhost:11434/api/chat -d '{
  "model": "zerodaylk/GHOSTPROTOCOL",
  "messages": [{"role": "user", "content": "Explain SSRF in two sentences."}],
  "stream": false
}'

import ollama

r = ollama.chat(model="zerodaylk/GHOSTPROTOCOL",
                messages=[{"role": "user", "content": "What is a SYN flood?"}])
print(r["message"]["content"])

Model card

Default parameters

Where it earns its place

• Offline and air-gapped labs — no API key, no egress, no telemetry.
• Implant and agent scaffolding — a local language backend small enough to ship inside tooling.
• CI and fuzzing harnesses — cheap enough to call in a loop.
• Teaching and CTF infrastructure — hand every participant their own model.
• Edge hardware — Pi-class devices, dropboxes, constrained VMs.

Where it does not

Be straight about the size. At 134M parameters GHOSTPROTOCOL is not a security knowledge oracle and must not be treated as one:

• It hallucinates freely. Never take a CVE number, a payload, a port, or a command from it without verifying.
• It has no reliable knowledge of specific vulnerabilities, tooling flags, or exploit details.
• It cannot do multi-step reasoning, long-context analysis, or code audit.
• 2048-token context. That is roughly three pages.

Use it for scaffolding, phrasing, classification, glue and offline convenience. Use a frontier model for anything that needs to be correct.

Responsible use

GHOSTPROTOCOL is intended for authorized security work only — scoped engagements with written permission, CTFs, detection engineering, and defensive research. Testing systems you do not own or have explicit authorization to test is illegal in most jurisdictions. You are responsible for your own scope.

Notes on the artifact

Two things are worth stating plainly rather than leaving for someone to find:

1. tokenizer.ggml.pre is unchanged. That key selects the pre-tokeniser regex inside the inference engine — it is a behaviour switch, not a label. Changing it was tested and measurably altered model output, so it stays.
2. Metadata branding is not provenance erasure. The weights are unmodified, so the upstream base model remains trivially identifiable by anyone who hashes a tensor or inspects the vocabulary. This package rebrands the artifact; it does not and cannot obscure its lineage.

Licence

Apache-2.0. The full licence text ships inside the model.

Built on Apache-2.0 licensed open weights. GHOSTPROTOCOL is a repackaging and configuration of those weights by zeroday.lk; the underlying model is unmodified and remains the work of its original authors.