A Different Approach to Cybersecurity AI
OpenAI’s GPT-6 Astra has demonstrated advanced cybersecurity capabilities, including a reported perfect score on ExploitBench in OpenAI’s published evaluations. Against that backdrop, Indian entrepreneur-scientist Fazilulla Baig Mirza is pursuing a different approach: exploring whether specialized cybersecurity AI can deliver strong results without relying on frontier-scale models and infrastructure. His work, including research associated with ExploitBench, focuses on efficient models, reproducible evaluation and practical cybersecurity applications.
From Security Research to AI Engineering
Fazilulla Baig Mirza, Founder and CEO of FAC INFOSEC PRIVATE LIMITED (FAC Security), works across cybersecurity research, artificial intelligence and machine learning. His background includes Web2 security, Web3 security, vulnerability research, bug bounty research, LLM security, AI security, exploit research and autonomous AI systems. This combination of offensive and defensive security experience informs his broader interest in how AI can change vulnerability discovery, security testing and cyber defense.
The GPT-6 Astra Challenge
The central question behind Fazilulla’s reported experiment is whether advanced cybersecurity performance necessarily requires the scale associated with frontier AI systems. His approach emphasizes specialized cybersecurity fine-tuning, efficient infrastructure and reproducible evaluation rather than simply increasing model size.
According to the results he reports, a 14-billion-parameter cybersecurity-specialized model was evaluated on ExploitBench and achieved 150/150 responses, or 100%, in his test. The result is presented as an experiment rather than as a general claim that a smaller model has surpassed GPT-6 Astra across cybersecurity. OpenAI separately reports that GPT-6 Astra achieved 100% on its ExploitBench evaluation, making the benchmark an important reference point for discussions about AI-assisted exploit development.
Why Reproducibility Matters
Beyond the headline score, the reported experiment places emphasis on deterministic generation, SHA-256 hashing and reproducible methodology. Such practices can help researchers examine whether an evaluation can be independently repeated and whether reported outcomes remain consistent under the same conditions.
ExploitBench itself is designed to measure how far AI agents progress through exploitation tasks, from reaching vulnerable code and triggering bugs to developing exploit primitives and achieving arbitrary code execution. The benchmark therefore provides a more technical evaluation environment than a general-purpose AI test.
Building Autonomous Cyber Defense
The research connects with Fazilulla’s wider work at FAC Security, including CyberDefenceX, described as an AI-based cyber defense architecture. Its stated focus includes threat detection, autonomous response and containment, adaptive deception and self-healing cybersecurity.
The broader objective is to move from security systems that primarily generate alerts toward systems that can support a continuous defensive loop: detecting suspicious activity, analysing threats, responding to incidents, assisting remediation and verifying security outcomes. Zero-day research, attack-path analysis and machine-speed response form part of this wider direction.
From Cybersecurity AI to AI Infrastructure
Fazilulla is also developing NazzBuild, an AI-native infrastructure concept focused on AI agents, autonomous workflows, application generation, organizational intelligence and AI infrastructure. Together with CyberDefenceX, the project reflects a broader interest in building systems where AI agents can perform increasingly complex operations while security is integrated into the underlying architecture.
For the cybersecurity industry, this raises a wider question: if AI-assisted attacks can become faster and more automated, how should defensive systems evolve to operate at comparable speed while maintaining appropriate safeguards?
A Broader Experiment in Efficient AI
The significance of Fazilulla’s reported work lies less in a single benchmark number and more in the engineering question it explores. Can specialized models, efficient infrastructure and rigorous evaluation make advanced cybersecurity AI more accessible?
GPT-6 Astra’s published cybersecurity results demonstrate the capabilities possible at the frontier, while Fazilulla’s reported experiments approach the same broader problem from a different direction. The continuing development of specialized cybersecurity models may offer researchers another path for testing how model size, domain specialization, infrastructure efficiency and evaluation methodology influence practical AI performance.
Backlink: ExploitBench research repository
