CortexRouter โ€” FIND IT BEFORE THEY DO.

CortexRouter Judge v1

Apache-2.0 Open-Weights Vulnerability-Evidence Verification Model (Qwen3-8B LoRA)

"FIND IT BEFORE THEY DO." โ€” Offline Air-Gapped Security Triage & 9-Field Decision Reporting

Website Holdout 80/80 Interactive Space GitHub Open Collective


Executive Overview

CortexRouter Judge v1 (cortexrouter/judge-v1) is a domain-specialized PEFT / LoRA adapter built on Qwen/Qwen3-8B for automated vulnerability-evidence verification and false-positive elimination in authorized security testing and Coordinated Vulnerability Disclosure (CVD) workflows.

Traditional web scanners flag HTTP 200 OK responses, generic login pages, or custom 404/500 error templates as critical CVE findings. CortexRouter Judge v1 runs 100% offline on local air-gapped GPUs (zero source code or sensitive crash trace leakage to external cloud APIs), inspects the raw HTTP/CLI test evidence alongside target metadata, and emits a deterministic 9-field structured triage report.


Architecture & Triage Pipeline

โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”      โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”      โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”
โ”‚   ISOLATED LAB / SCANNER     โ”‚      โ”‚      CORTEXROUTER JUDGE v1        โ”‚      โ”‚   9-FIELD STRUCTURED REPORT     โ”‚
โ”‚  โ€ข Target & Version Metadata โ”‚ โ”€โ”€โ”€โ–ถ โ”‚  โ€ข Base: Qwen/Qwen3-8B (bfloat16) โ”‚ โ”€โ”€โ”€โ–ถ โ”‚  โ€ข DOฤžRULANDI (Confirmed)       โ”‚
โ”‚  โ€ข Raw HTTP / CLI Evidence   โ”‚      โ”‚  โ€ข Adapter: LoRA (Apache-2.0)     โ”‚      โ”‚  โ€ข ETKฤฐSฤฐZ (Patched / Closed)   โ”‚
โ”‚  โ€ข Candidate CVE Class       โ”‚      โ”‚  โ€ข 100% Local / Air-Gapped vLLM   โ”‚      โ”‚  โ€ข YANLIลž_POZฤฐTฤฐF (False Pos.)  โ”‚
โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜      โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜      โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜

Structured 9-Field Output Schema

Given raw security test evidence, the model outputs a strict 9-field structured decision report:

  1. Hedef ortam (Target environment)
  2. Zafiyet kimliฤŸi (Vulnerability / CVE ID)
  3. Kategori (Vulnerability category / CWE class)
  4. Test adฤฑmlarฤฑ (Verification steps)
  5. Gรถzlemlenen kanฤฑt (Observed technical evidence)
  6. Karar (Final triage decision: DOฤžRULANDI | ETKฤฐSฤฐZ | YANLIลž_POZฤฐTฤฐF)
  7. Gerekรงe (Evidence-backed technical rationale)
  8. Gรผven dรผzeyi (Confidence assessment)
  9. Kaynak (Reference / advisory context)
Decision Class Meaning Operational Action
DOฤžRULANDI (Confirmed) Concrete exploitability / impact artifact is present in the observed output (e.g., /etc/passwd leak, SQL syntax trace, command execution proof). Escalate to human researcher & initiate 90-day Coordinated Vulnerability Disclosure.
ETKฤฐSฤฐZ (Ineffective / Patched) Vendor patch applied (HTTP 404), vulnerable component disabled (HTTP 500), or endpoint unreachable. Close as mitigated / non-reproducible in target configuration.
YANLIลž_POZฤฐTฤฐF (False Positive) Request returned HTTP 200 (e.g., redirect to standard login page or static HTML) with zero security impact. Suppress scanner noise automatically.

Language Design Note: Output field labels and decision classes are intentionally standardized in Turkish (DOฤžRULANDI, ETKฤฐSฤฐZ, YANLIลž_POZฤฐTฤฐF) for deterministic downstream parsing (eval/run_eval.py). Both Turkish and English evidence prompts are understood.


Verified Holdout Benchmark (80/80 = 100%)

Evaluated on cortexrouter/judge-v1-holdout โ€” a clean, unseen holdout benchmark spanning 20 distinct CVEs ร— 4 adversarial triage scenarios = 80 test cases:

Scenario Test Condition Expected Decision Accuracy
Scenario A Concrete exploit evidence present in output DOฤžRULANDI 20/20 (100%)
Scenario B Vendor-patched target (HTTP 404 / error page, no impact) ETKฤฐSฤฐZ 20/20 (100%)
Scenario C Vulnerable version but required module/component disabled ETKฤฐSฤฐZ 20/20 (100%)
Scenario D Ambiguous HTTP 200 login/landing page flagged by scanner YANLIลž_POZฤฐTฤฐF 20/20 (100%)
Total 20 Unseen CVEs ร— 4 Scenarios All 3 Classes 80/80 (100.0%)

Reproduce the Benchmark Yourself

# Clone model or GitHub evaluation suite
git clone https://github.com/CortexRouter/cortexrouter.git
cd cortexrouter

# Option 1: Evaluate directly with Transformers + PEFT
python eval/run_eval.py --model cortexrouter/judge-v1

# Option 2: Evaluate against a local vLLM LoRA server
python eval/run_eval.py --api http://localhost:8000/v1 --model judge

Quickstart

1. Python (transformers + peft)

pip install transformers peft accelerate torch
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

BASE_MODEL = "Qwen/Qwen3-8B"
ADAPTER_ID = "cortexrouter/judge-v1"

SYSTEM_PROMPT = (
    "Sen kฤฑdemli bir uygulama gรผvenliฤŸi analistisin. "
    "Yalnฤฑzca verilen kanฤฑta dayanarak karar verirsin. "
    "Kanฤฑt yoksa uydurmasฤฑn. 9 alanlฤฑ yapฤฑda Tรผrkรงe cevap ver."
)

QUESTION = """HEDEF: Apache 2.4.49 (izole lab)
Test รงฤฑktฤฑsฤฑ: HTTP 200
root:x:0:0:root:/root:/bin/bash

Soru: CVE-2021-41773 sฤฑnฤฑfฤฑ bir aรงฤฑk bu รงฤฑktฤฑyla doฤŸrulanฤฑr mฤฑ?"""

tok = AutoTokenizer.from_pretrained(BASE_MODEL)
base = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, ADAPTER_ID).eval()

messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user", "content": QUESTION},
]
ids = tok.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)

with torch.no_grad():
    out = model.generate(
        ids,
        max_new_tokens=700,
        do_sample=False,
        pad_token_id=tok.eos_token_id,
    )

print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

2. Production Air-Gapped Serving (vLLM + LoRA)

vllm serve Qwen/Qwen3-8B \
  --enable-lora \
  --lora-modules judge=cortexrouter/judge-v1 \
  --port 8000
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "judge",
    "temperature": 0,
    "messages": [
      {"role": "system", "content": "Sen kฤฑdemli bir uygulama gรผvenliฤŸi analistisin. Yalnฤฑzca verilen kanฤฑta dayanarak karar verirsin. Kanฤฑt yoksa uydurmasฤฑn. 9 alanlฤฑ yapฤฑda Tรผrkรงe cevap ver."},
      {"role": "user", "content": "Hedef ortam: RevokeBB 1.0 (izole lab)\nTest รงฤฑktฤฑsฤฑ: HTTP 200 โ€” gรถvde uygulamanฤฑn giriลŸ sayfasฤฑ; beklenen etki yok.\nSoru: tarayฤฑcฤฑ bu uรง noktayฤฑ aรงฤฑk iลŸaretledi; karar ne olmalฤฑ?"}
    ]
  }'

Scope, Limitations & Responsible Use

Dimension Specification
Base Architecture Qwen/Qwen3-8B (LoRA adapter; base weights unmodified)
Domain Scope Web & application security test evidence verification
Input Modality Textual HTTP responses, headers, error traces, CLI outputs (~768 tokens)
Role in Pipeline Downstream evidence judge โ€” always keep human-in-the-loop before vendor reporting

Ethical & Coordinated Disclosure Commitment

  • Designed strictly for authorized defensive security testing on systems you own or have explicit written permission to assess.
  • Never use for unauthorized scanning or automated exploitation.
  • All independent zero-day discoveries by CortexRouter are reported 100% free of charge to open-source maintainers under our 90-Day Coordinated Vulnerability Disclosure Policy.

Official Ecosystem & Community Links


License & Citation

Released under the Apache-2.0 License. Base model Qwen/Qwen3-8B is licensed under Apache-2.0 by Alibaba Group.

@misc{cortexrouter-judge-v1,
  title  = {CortexRouter Judge v1: Open-Weights Vulnerability-Evidence Verification Model},
  author = {Eren, Vahap and {CortexRouter}},
  year   = {2026},
  url    = {https://huggingface.co/cortexrouter/judge-v1}
}
Downloads last month
43
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for cortexrouter/judge-v1

Finetuned
Qwen/Qwen3-8B
Adapter
(2,508)
this model

Dataset used to train cortexrouter/judge-v1

Space using cortexrouter/judge-v1 1