Rowhanissick/Verfy-9B

Developed by Rowhanissick & Verfy Labs
Base Architecture: empero-ai/Qwythos-9B-v2 (Qwen3.5 9B Hybrid-Attention)
Lead Engineer: Rohan Patil (Rowhanissick)
Portfolio: rowhanissick.onrender.com
Direct Contact: iuske553@gmail.com

---

🌟 Overview

Verfy AI v1 is an advanced autonomous OS-agent, cybersecurity reasoning engine, and self-auditing intelligence model engineered for desktop/server automation, red-teaming, threat intelligence, and secure execution.

Unlike standard conversational chatbots that only output unstructured text, Verfy natively reasons through its operational stream, executes self-auditing reflection passes (<big_bro_audit>), and isolates sensitive credentials before executing environment actions (CLI, mouse, keyboard, APIs).

---

🛡️ Key Breakthrough Capabilities

1. Big Bro Self-Reflection (<big_bro_audit>)

Before finalizing an execution output or responding to ambiguous prompts, Verfy performs an internal cognitive pass inside <big_bro_audit> tags:

• Critiques logical consistency and command safety.
• Catches potential hallucinated mock scripts before deployment.
• Verifies authorization boundaries and prevents social engineering traps.

2. Autonomous Agentic Safeguards

• Credential Enclave Awareness: Recognizes PII (emails, phone numbers, secret tokens) and flags them for secure vaulting rather than echoing in plain text.
• Command-Gated Execution: Refuses voluntary unauthorized broadcasting (social media, public outbound traffic) without explicit operator authorization tokens.

3. Native Block Architecture

Verfy AI communicates and executes tasks using structured XML-style tags:

Block Tag	Purpose	Example
<think>	Deep internal reasoning & strategy planning	Formulating execution path and identifying risks
<big_bro_audit>	Pre-output self-verification & logic check	Auditing generated script for syntax & logic bugs
<steps>	Decomposes operations into sequential tasks with <title>	1. Locate browser 2. Dispatch click
<checks>	Precondition validation & safety checks with <title>	- Screen bounds verified<br>- Non-destructive check passed
<screenshot>	Requests visual frame capture of screen/window	capture(mode='screen', region='desktop')
<command>	Shell / Bash terminal command execution	ss -tulnp | grep 8080
<read>	Safe file, log, or memory dump inspection	read_file(path='/var/log/syslog', lines=50)
<write>	File patching and exploit/payload deployment	write_file(path='/tmp/patch.py', content='...')
<input>	GUI mouse movements and keystroke dispatches	mouse_move(x=500, y=500) mouse_click('left')
<vuln>	Flags identified CVEs & security vulnerabilities	ID: VULN-001 Severity: HIGH CWE-78

---

⚡ Headline Specifications

• 1,048,576 Token Context: Full 1M-token context window out of the box via YaRN rope-scaling, allowing whole-codebase and long agentic trajectory tracking.
• Uncensored Technical Reasoning: Built on an abliterated Qwen 3.5 architecture (Qwythos-9B-v2), Verfy provides unredacted technical insights for cybersecurity research, reverse engineering, and threat modeling.
• Dual Interaction Modes: Operates both at the conversational level and the autonomous agent level (<command> + <input> + <screenshot>).

---

💻 How to Use (Python)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

MODEL_ID = "Rowhanissick/Verfy-9B"

quant_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16
)

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    quantization_config=quant_config,
    device_map="auto",
    torch_dtype=torch.float16,
    trust_remote_code=True
)

prompt = "Analyze open listening sockets and report external exposure."
messages = [{"role": "user", "content": prompt}]

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=512,
    temperature=0.6,
    top_p=0.9
)

print(tokenizer.decode(outputs[0][len(inputs.input_ids[0]):], skip_special_tokens=True))

---

📦 GGUF & Local Deployments

Quantized versions (Q4_K_M, Q8_0) compatible with llama.cpp, Ollama, and LM Studio are available via community builds:

• mradermacher/Verfy-9B-GGUF

---

👤 Author & Organization

• Organization: Verfy Labs
• Lead Developer: Rohan Patil (Rowhanissick)
• Portfolio: rowhanissick.onrender.com
• Inquiries & Security Reports: iuske553@gmail.com
• Hugging Face Profile: @Rowhanissick
• License: Apache 2.0 (Open-source, free for research, red-teaming, and autonomous development)