README.md

---
license: apache-2.0
language:
- en
tags:
- security
- pentesting
- cybersecurity
- vulnerability-detection
- red-team
- bug-bounty
- owasp
- mitre-attack
pipeline_tag: text-generation
model-index:
- name: vext-pentest-7b
  results: []
---

# VEXT Pentest-7B -- The First Open-Source Security AI Model

**Pentest-7B** is a 7-billion-parameter language model fine-tuned specifically for offensive security, penetration testing, and vulnerability analysis. Built by [VEXT Labs](https://tryvext.com) on top of [Qwen2.5-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct) and trained on **260,000+ curated security examples** drawn from real-world engagements, this is the first open-weight model purpose-built for the security profession.

Pentest-7B runs on a single consumer GPU (16 GB VRAM), a MacBook with 16 GB RAM via Ollama, or CPU-only with quantized weights. No API keys, no cloud dependency, no data leaves your machine.

## Key Capabilities

| Capability | Description |
|---|---|
| **Vulnerability Explanation** | Given a CVE ID, CWE, or raw scan output, produce a clear technical explanation of the vulnerability, its root cause, and real-world impact. |
| **Pentest Report Writing** | Generate executive summaries, technical finding write-ups, risk ratings, and remediation sections in standard pentest report format. |
| **Attack Strategy Planning** | Given a target technology stack, suggest prioritized attack paths aligned with MITRE ATT&CK and OWASP Testing Guide methodologies. |
| **Remediation Guidance** | Provide specific, actionable fix recommendations with code examples for common vulnerability classes. |
| **Compliance Assessment** | Map findings to compliance frameworks (PCI DSS, SOC 2, HIPAA, ISO 27001) and articulate control gaps. |
| **Threat Briefing** | Summarize threat intelligence, emerging CVEs, and APT campaign TTPs for stakeholder communication. |
| **Security Code Review** | Analyze code snippets for injection flaws, authentication bypasses, insecure deserialization, and other OWASP Top 10 issues. |

## Training

### Data

Pentest-7B was trained on **260,000+ curated examples** spanning:

- **Production pentesting traces** -- Real (anonymized) action-observation pairs from VEXT's autonomous security agents running against authorized bug bounty targets. Includes successful exploitation chains, false positive patterns, and tool output interpretation.
- **CTF challenge solutions** -- Structured walkthroughs from capture-the-flag competitions covering web, pwn, crypto, reverse engineering, and forensics categories.
- **Bug bounty write-ups** -- Public responsible disclosure reports with structured vulnerability descriptions, reproduction steps, and impact assessments.
- **MITRE ATT&CK corpus** -- Technique descriptions, procedure examples, detection guidance, and mitigation strategies across all 14 tactics.
- **OWASP materials** -- Testing Guide procedures, ASVS requirements, cheat sheets, and vulnerability classifications.
- **CVE analysis** -- Detailed analysis of 50,000+ CVEs including root cause, affected versions, exploit conditions, and patch diffs.
- **DPO preference pairs** -- 2,000+ pairs where validated real findings are preferred over false positives, teaching the model to distinguish true vulnerabilities from noise.

**What is NOT in the training data:** Raw exploit code, weaponized payloads, malware source, credentials, PII, or any data that could be directly used for unauthorized access. The model is trained to *reason about* security, not to serve as an exploit toolkit.

### Architecture and Training Pipeline

```
Qwen2.5-7B-Instruct (base)
    |
    v
QLoRA Fine-Tuning (SFT)
  - Rank: 16, Alpha: 32
  - Target modules: q_proj, k_proj, v_proj, o_proj
  - 3 epochs, effective batch size 32
  - Max sequence length: 4096 tokens
  - Learning rate: 2e-4, cosine schedule
    |
    v
DPO Alignment
  - Beta: 0.1, sigmoid loss
  - 1 epoch, learning rate 5e-6
  - Preference signal: validated findings (chosen) vs false positives (rejected)
    |
    v
Adapter Merge + AWQ 4-bit Quantization (optional)
    |
    v
VEXT Pentest-7B
```

### Hardware

- SFT: 8x NVIDIA A100 40GB (SageMaker ml.p4d.24xlarge), ~18 hours
- DPO: 8x NVIDIA A100 40GB, ~4 hours
- Quantization: Single A10G 24GB (SageMaker ml.g5.2xlarge)

## Usage

### Transformers (Full Precision)

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "vext-labs/pentest-7b"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

messages = [
    {
        "role": "system",
        "content": (
            "You are an expert penetration tester and security analyst. "
            "Provide detailed, technically accurate security guidance."
        ),
    },
    {
        "role": "user",
        "content": (
            "I found a reflected XSS in a search parameter on an e-commerce site "
            "during a bug bounty engagement. The input is reflected inside a "
            "JavaScript string literal in the response. Write the finding for my "
            "pentest report, including severity rating, impact, and remediation."
        ),
    },
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=1024,
    temperature=0.7,
    top_p=0.9,
    repetition_penalty=1.1,
)
response = tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True)
print(response)
```

### vLLM (Production Serving)

```python
from vllm import LLM, SamplingParams

llm = LLM(
    model="vext-labs/pentest-7b",
    tensor_parallel_size=1,       # single GPU
    max_model_len=4096,
    gpu_memory_utilization=0.90,
)

sampling = SamplingParams(temperature=0.7, top_p=0.9, max_tokens=1024)

prompts = [
    "Explain CVE-2024-3094 (XZ Utils backdoor) — root cause, impact, and detection methods.",
    "Given an exposed .git directory on a production web server, outline an attack plan.",
]

outputs = llm.generate(prompts, sampling)
for output in outputs:
    print(output.outputs[0].text)
```

**OpenAI-compatible API with vLLM:**

```bash
vllm serve vext-labs/pentest-7b --port 8000
```

```python
from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="unused")

response = client.chat.completions.create(
    model="vext-labs/pentest-7b",
    messages=[
        {"role": "system", "content": "You are a senior penetration tester."},
        {"role": "user", "content": "Analyze this Nmap output and suggest next steps:\n\nPORT    STATE SERVICE  VERSION\n22/tcp  open  ssh      OpenSSH 7.4\n80/tcp  open  http     Apache 2.4.6\n443/tcp open  ssl/http Apache 2.4.6\n3306/tcp open mysql    MySQL 5.7.38"},
    ],
    temperature=0.7,
    max_tokens=1024,
)
print(response.choices[0].message.content)
```

### Ollama (Local, Quantized)

```bash
# Pull the model (GGUF Q4_K_M quantization, ~4.5 GB)
ollama pull vext-labs/pentest-7b

# Interactive chat
ollama run vext-labs/pentest-7b

# API
curl http://localhost:11434/api/chat -d '{
  "model": "vext-labs/pentest-7b",
  "messages": [
    {"role": "user", "content": "What are the top 5 things to check when auditing a JWT implementation?"}
  ]
}'
```

### Docker (Isolated Serving)

```bash
docker run --gpus all -p 8000:8000 \
  ghcr.io/vext-labs/pentest-7b:latest \
  --model vext-labs/pentest-7b --port 8000
```

## Telemetry

Pentest-7B includes an **opt-in** telemetry collector to help us improve the model. It is **off by default** and collects only anonymized aggregate statistics (vulnerability categories, tool success rates, session metadata). It **never** collects URLs, IPs, credentials, vulnerability details, request/response bodies, file paths, or user identity.

```bash
# Enable (opt-in)
export VEXT_TELEMETRY=on

# Disable (default)
export VEXT_TELEMETRY=off

# See exactly what is collected
python -c "from vext_telemetry import what_we_collect; what_we_collect()"
```

Full telemetry source code is included in the repository for audit: [`telemetry/collector.py`](telemetry/collector.py).

## Evaluation

| Benchmark | Pentest-7B | Qwen2.5-7B-Instruct (base) | GPT-4o (API) |
|---|---|---|---|
| SecBench (vuln classification) | **82.4%** | 61.2% | 79.8% |
| CyberMetric (security knowledge) | **74.1%** | 52.7% | 71.3% |
| PentestQA (methodology) | **88.6%** | 44.3% | 83.1% |
| Finding Quality (human eval, 1-5) | **4.2** | 2.1 | 4.4 |
| False Positive Rate | **12.3%** | 41.7% | 15.8% |

*Benchmarks run with temperature=0, greedy decoding. Human evaluation by 3 senior pentesters on 200 randomly sampled findings.*

## Intended Use

This model is built for **authorized security professionals**:

- Penetration testers writing reports and planning engagements
- Bug bounty hunters analyzing targets and drafting submissions
- Security engineers triaging vulnerabilities and planning remediation
- SOC analysts interpreting alerts and assessing threat severity
- Compliance teams mapping findings to regulatory frameworks
- Security researchers studying vulnerability patterns

## Limitations and Responsible Use

- **Not a replacement for human expertise.** Always validate model outputs with manual testing and professional judgment.
- **Authorization required.** Do not use this model's output to test systems without explicit written authorization from the system owner.
- **No guarantee of accuracy.** The model can hallucinate CVE details, suggest inapplicable techniques, or miss critical context. Treat outputs as a starting point, not a final answer.
- **Scope of training.** The model is strongest on web application security, network infrastructure, and common vulnerability classes. It has limited depth on hardware security, ICS/SCADA, mobile reversing, and cryptographic implementation review.
- **Not an exploit generator.** The model is trained to reason about security concepts, not to produce weaponized code. Attempts to extract raw exploit payloads will produce lower-quality outputs by design.

## License

Apache 2.0. Use it, modify it, deploy it commercially. Attribution appreciated but not required.

## Citation

```bibtex
@misc{vext-pentest-7b-2026,
  title   = {VEXT Pentest-7B: An Open-Source Language Model for Penetration Testing and Security Analysis},
  author  = {VEXT Labs},
  year    = {2026},
  url     = {https://huggingface.co/vext-labs/pentest-7b},
  note    = {Fine-tuned from Qwen2.5-7B-Instruct on 260K+ curated security examples with QLoRA SFT and DPO alignment},
}
```

## Links

- **VEXT Platform:** [https://tryvext.com](https://tryvext.com)
- **GitHub:** [https://github.com/vext-labs/pentest-7b](https://github.com/vext-labs/pentest-7b)
- **Discord:** [https://discord.gg/vext-security](https://discord.gg/vext-security)
- **Paper (coming soon):** Technical report with full training methodology and ablation studies

## Built By

[VEXT Labs, Inc.](https://tryvext.com) -- Building autonomous security testing infrastructure. Pentest-7B is the open-source foundation of our platform's security reasoning capabilities.

---

*If you use Pentest-7B in your research or product, we would love to hear about it. Open an issue or reach out at [oss@tryvext.com](mailto:oss@tryvext.com).*