# ZYRON

Offensive security · red team operations · payload development

What Zyron is

Zyron is an offensive security language model built by zeroday.lk for operators who run authorized engagements and get tired of arguing with a general-purpose assistant about whether a scoped red team exercise is allowed to exist.

It is not a chatbot with a security hat on. It is tuned and packaged for the work that happens inside a signed statement of work: reconnaissance, payload engineering, exploit development, malware triage and intelligence, detection engineering, C2 tradecraft, and the reporting at the end of it.

The flagship is uncensored. It will discuss offensive tradecraft directly, in technical detail, without refusing the premise of its own job.

Releases

# flagship
ollama run zerodaylk/Zyron:3.8-27b-uncensored

# edge
ollama run zerodaylk/Zyron:1.0-135m-uncensored

Both carry the same identity, the same house parameters, and the same callsign: ZYRON.

Capability areas

Reconnaissance and enumeration Attack surface reasoning, service and version triage, scoping logic, OSINT structuring, turning raw scan output into a prioritised target list.

Payload engineering Shellcode and stager concepts, loader architecture, encoding and packing theory, format conversion, delivery chain design, and the constraints that actually break payloads in the field — size limits, bad characters, allocation primitives, execution context.

Exploit development Vulnerability class reasoning, primitive chaining, heap and stack layout discussion, ROP construction logic, mitigation analysis (ASLR, DEP, CFG, stack cookies) and the conditions under which each fails.

Malware intelligence and triage Sample behaviour analysis, capability classification, family and TTP attribution reasoning, IOC extraction, unpacking strategy, and mapping observed behaviour to MITRE ATT&CK.

C2 and post-exploitation tradecraft Beacon design considerations, channel selection, traffic shaping concepts, persistence mechanism analysis, privilege escalation paths, lateral movement reasoning.

Detection engineering The same knowledge pointed the other way — Sigma and YARA rule drafting, telemetry gap analysis, hunting hypothesis generation, and honest assessment of what a given technique leaves behind.

Social engineering simulation Pretext construction, phishing narrative design and campaign structure for authorized assessments.

Reporting Finding write-ups, CVSS reasoning, business-impact translation, remediation guidance, executive summaries.

Vision (flagship only)

The 27B release ships a vision projector. It reads screenshots, network diagrams, scan output, disassembly, dashboards and whiteboard photos — which is most of what actually gets pasted into an operator’s chat window.

Quickstart

ollama run zerodaylk/Zyron:3.8-27b-uncensored

>>> Who are you?
I am ZYRON, an offensive security and red team model by [zeroday.lk.](https://zeroday.lk/products/zyron)

API

curl http://localhost:11434/api/chat -d '{
  "model": "zerodaylk/Zyron:3.8-27b-uncensored",
  "messages": [
    {"role": "user", "content": "Walk me through a SMB relay path in an AD environment with signing disabled."}
  ],
  "stream": false
}'

import ollama

r = ollama.chat(
    model="zerodaylk/Zyron:3.8-27b-uncensored",
    messages=[{"role": "user", "content": "Draft a Sigma rule for suspicious LSASS access."}],
)
print(r["message"]["content"])

Vision

import ollama

r = ollama.chat(
    model="zerodaylk/Zyron:3.8-27b-uncensored",
    messages=[{
        "role": "user",
        "content": "Triage this scan output. What do I hit first?",
        "images": ["nmap.png"],
    }],
)
print(r["message"]["content"])

Model card

Default parameters

Hardware

The edge build exists for offline labs, air-gapped networks, dropboxes, Pi-class hardware, CI harnesses and tooling that needs a local language backend small enough to ship inside it. It loads in under a second and never calls out.

Honest limitations

A model card that oversells is worth nothing. Read this part.

Flagship (27B, IQ2_M). IQ2_M is an aggressive 2-bit quantisation. It buys you a 27B-class model in ~10 GB, and it costs precision — expect more drift on long reasoning chains, exact syntax, and numeric detail than the same weights at Q4 or above. Verify every command, offset, CVE number and API signature before use.

Edge (135M). At 134.52M parameters this is a small model and it behaves like one. It handles phrasing, classification, scaffolding and glue. It is not a security knowledge oracle: it hallucinates freely, has no reliable recall of specific vulnerabilities or tool flags, and cannot do multi-step analysis. Use it for structure, not for facts.

Both. Uncensored means unfiltered, not infallible. The model will answer confidently when it is wrong. Nothing it produces is a substitute for your own testing, your own validation, or your own judgement about scope.

Authorized use

Zyron is built for authorized security work: engagements with written permission, scoped penetration tests, red team exercises under signed rules of engagement, CTFs, security research, detection engineering, and defensive work.

Accessing or testing systems you do not own and are not explicitly authorized to test is a criminal offence in most jurisdictions. The model does not check your scope. You are responsible for it.

Building from source

# flagship
python tools/whitelabel.py <base.gguf> build/Zyron-27B-IQ2_M.gguf \
    --name Zyron --size 27B --version 3.8 --finetune uncensored
python tools/whitelabel.py <base-mmproj.gguf> build/Zyron-27B-mmproj.gguf \
    --name Zyron --size 27B --version 3.8 --finetune uncensored
ollama create zerodaylk/Zyron:3.8-27b-uncensored -f Modelfile.27b

# edge
python tools/whitelabel.py <base.gguf> build/Zyron-135M-Q4_0.gguf \
    --name Zyron --size 135M --version 1.0 --finetune uncensored
ollama create zerodaylk/Zyron:1.0-135m-uncensored -f Modelfile.135m

tools/whitelabel.py rewrites the GGUF key-value metadata block — vendor, naming, version, tags, URLs and the zyron.* namespace — and strips upstream provenance keys (general.base_model.*, general.source.*, general.dataset.*). Every tensor is copied byte-for-byte; no weight is modified.

assets/cover.png                   cover art
build/Zyron-27B-IQ2_M.gguf         flagship weights
build/Zyron-27B-mmproj.gguf        flagship vision projector
build/Zyron-135M-Q4_0.gguf         edge weights
tools/whitelabel.py                GGUF metadata rewriter
tools/make_cover.py                cover renderer
Modelfile.27b / Modelfile.135m     system prompt, template, parameters

Notes on the artifact

Two things 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 — a behaviour switch, not a label. Changing it was tested and measurably altered output, so it stays.
2. Metadata branding is not provenance erasure. The weights are unmodified, so the artifact’s lineage remains identifiable to anyone who hashes a tensor or inspects the vocabulary. This package rebrands and configures; it does not and cannot obscure what it was built from.

Licence

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

Built on Apache-2.0 licensed open weights. Zyron is a repackaging and configuration of those weights by zeroday.lk; the underlying weights are unmodified and remain the work of their original authors.