Abanisenioluwa "Banny" Orojo

Computer Science (Cybersecurity) - CybearSecurity Lab Central Texas Cyber Range · 
Email: Abanisenioluwa_Oroj1@baylor.edu Website: labackdoor.com · GitHub: github.com/LaBackDoor

Who I Am

I'm a PhD candidate in Computer Science at Baylor, where I lead research in the CybearSecurity Lab on data-driven defense for cyber-physical and networked systems. My work applies machine learning and language models to detection, access control, protocol analysis, and automated repair, and I test all of it against live, running systems rather than static datasets.

What I'm Working On

Autonomous vulnerability remediation. My core research asks whether AI agents can find and fix real vulnerabilities on live hosts without breaking the services running on them. That last part is the hard constraint: a patch that closes a security hole but takes down the system it was protecting isn't a fix; it's a different failure. Most real-world remediations are multi-step and order-dependent get the sequence wrong (say, enabling a firewall before adding the right allow-rule), and you lock yourself out of the system you're trying to secure.

To study this properly, I built SysRepair-Bench, a benchmark of 313 scenarios spanning two decades of real-world threats from modern CVEs like Log4Shell and PwnKit to legacy systems where patching isn't even an option and the agent has to find a workaround. I then evaluated a range of AI agent architectures against it and developed NeuroPlan, an approach that translates a compromised host into a formal planning problem so a human operator can inspect a plan for consistency before any change is made to the system rather than just trusting the agent's judgment.

Two findings from this work reshape how I think about deploying these agents in practice:

• Context matters more than reasoning ability. When agents aren't told which vulnerability to look for, accuracy collapses by roughly two-thirds. Scanner findings, CVE feeds, and asset inventories turn out to matter as much as which AI architecture you use.
• Today's agents triage; they don't close. On hosts with multiple simultaneous vulnerabilities, even the best-performing model only fully cleared 1 of 16 hosts, despite reporting over 50% of issues mitigated. That gap between "partial credit" and "actually secure" is exactly the kind of thing that needs to be surfaced before organizations rely on this technology.

Beyond remediation, my broader research spans:

• AI security- stress-testing learned models when applied to security tasks
• Threat detection & cyber-physical systems resilience- dynamic detection and access control for CPS environments
• Network & protocol intelligence- byte-level deep packet inspection using learned representations
• Security analytics- forecasting vulnerability trends from large, noisy data sources

Why This Matters

Security teams are increasingly being asked to trust AI agents with real remediation work, but most existing benchmarks reward finding an exploit or passing a test suite, not keeping a system alive. My research builds the evaluation infrastructure and planning methods needed to close that gap, so that autonomous defense tools can be deployed with a clear, measured understanding of what they can (and can't yet) be trusted to do.

Selected Publications

Under review

• Autonomous System Vulnerability Remediation: A Survey of Agentic AI, Reinforcement Learning, Benchmarks, and Operational Safety- Int. Journal of Information Security
• SysRepair-Bench: A Benchmark for AI Agents' Ability to Remediate Real-World System Vulnerabilities- ICLR 2027
• SysRepair: A Benchmark and Multi-Architecture Approach to Autonomous Vulnerability Remediation- ACSAC 2026

Published

• A Unified Framework Incorporating AW-TRBAC and Semantic Variational Autoencoders for Dynamic Threat Detection and Access Control- Int. Conference on Artificial Intelligence, 2025
• ByteFlow: A Byte-Level LLM for Deep Packet Inspection and Network Intelligence- ACSAC 2025 Workshops
• Assessing the Impact of Access Control Policies on Data Accessibility in Distributed NoSQL Environments- 24th Int. Conference on Security & Management, 2024
• Predicting Software Vulnerability Trends with Multi-Recurrent Neural Networks: A Time Series Forecasting Approach- First Int. Conference on NLP & AI for Cybersecurity, 2024
• Developing a Deep Learning Model for Detecting Cyber Attack, 2023
• Navigating the Digital Maze: Exploring the Intersection of Social Media, Privacy, and Security, 2022

Full list and code: labackdoor.com/publications