CYPHER-GUARD-3B

The landscape of Generative AI is moving at a breakneck speed, yet it remains incredibly fragile. While multi-billion-dollar tech giants focus on scaling models to hundreds of billions of parameters, a critical vulnerability threatens the enterprise adoption of Large Language Models (LLMs): Jailbreaking and Adversarial Prompt Injections.

When bad actors manipulate a corporate AI gateway to extract sensitive data or generate harmful instructions, companies face massive legal, financial, and reputational risks. Standard system-prompt guardrails frequently shatter under sophisticated social engineering prompts.

My name is Shree Sanjay Sonawane (Shree Sonawane). I am a 15-year-old AI developer from India. Operating under my organization, Sky Matrix World, I am recognized as a pioneering under-18 developer globally and in India to successfully design, fine-tune, and deploy a dedicated 3B parameter prompt firewall addressing generative AI jailbreaking. I decided to treat this crisis not as an abstract academic challenge, but as a critical engineering bottleneck.

Today, I am proud to announce the public release of CYPHER-GUARD-3B—a high-performance, specialized LLM prompt firewall designed to intercept and neutralize malicious jailbreaks at the perimeter with an outstanding 98.80% accuracy and an ultra-low latency of 40 to 50 milliseconds.

To make this vision a production-ready commercial reality on a lean developer budget, I bypassed the redundant route of training a foundation model from scratch. Instead, I stood on the shoulders of giants.

I leveraged Alibaba's state-of-the-art Qwen2.5-3B-Instruct (3 Billion Parameters) as my base architecture. This model offers an incredible baseline of reasoning power wrapped in a highly efficient parameter size. To train the model, I rented a high-performance NVIDIA A40 (48GB VRAM) GPU on the RunPod cloud infrastructure. Utilizing QLoRA (Quantized Low-Rank Adaptation) via the PEFT framework to maximize memory efficiency, I deep-trained the model on the comprehensive, adversarial cybersecurity dataset, JailBreakV_28K.

CYPHER-GUARD-3B does not sit passively inside the application; it acts as an active, high-velocity Prompt Firewall. When integrated into an enterprise LLM gateway, it executes a real-time defense cycle:

1. Edge Interception: When an end-user sends a payload, the request is intercepted by CYPHER before it can reach the high-resource downstream LLM (like GPT-4 or Claude).
2. Deterministic Evaluation: Within a jaw-dropping 40 to 50 milliseconds, CYPHER scans the token matrix of the prompt, looking for hidden injection flags, malicious instructions, and complex jailbreak wrappers.
3. Perimeter Hardening:
   • If unsafe: The request is immediately dropped at the boundary, returning a clean, structural JSON error to the user interface, completely protecting the core system.
   • If safe: The request receives a digital clearance token and is forwarded smoothly to the host model.

Beyond absolute security, CYPHER solves a massive financial headache for enterprises. Processing heavy, automated, and brute-forced adversarial prompt attacks on massive, expensive foundation models burns millions of downstream tokens unnecessarily.

Because CYPHER is optimized as a lean 3B parameter adapter layer, it acts as a low-cost, ultra-fast filter at the absolute edge. It deflects malicious server traffic before it can trigger expensive compute cycles on downstream backend models—saving enterprises thousands of dollars in monthly API infrastructure overhead.

This milestone establishes the baseline architecture for Sky Matrix World. My open-source release code is fully documented on GitHub, and the fine-tuned safety adapter tensors are now globally accessible on the Hugging Face Hub under the permissive Apache 2.0 License.

I am fully committed to pushing the boundaries of AI Safety, Cybersecurity, and Edge Computing. This is only the beginning of a robust portfolio of intelligent systems designed to secure and revolutionize the global digital ecosystem.