SysRepair-Bench
313 Dockerized and VM scenarios scored on two objectives: the exploit is closed and the service still serves. Six agent architectures evaluated.
Open to Summer 2027 Internships · Detection & Response, AI Security
Abanisenioluwa Orojo
PhD Candidate, Baylor University · Security & AI Researcher
I build and evaluate AI agents that find and fix vulnerabilities on live systems, and the cyber ranges, benchmarks and detection tooling around them. Creator of SysRepair-Bench, 313 reproducible attack-and-repair scenarios.
313
Scenarios in SysRepair-Bench
Multi-Site
GPU Training for RL Self-Play
10+
Papers and Preprints
2nd
WaCode 2026 · SWCCDC 2025
$ ls research/
Selected Research
NeuroPlan
Lifts a live host into PDDL via osquery and runs a precondition-checked plan an operator can review first. 76% end-to-end remediation.
ByteFlow
A byte-level LLM for deep packet inspection and network traffic intelligence.
Red/Blue Self-Play RL
Asynchronous attacker/defender training on a multi-site GPU cluster.
$ cat awards.log
Awards
2nd Place
WaCode 2026 Hackathon, 2nd Place
Team Barney and Friends placed second at Baylor's WaCode 2026 hackathon with Kyerɛ, a study app that makes students prove what they know by teaching an AI student. Orojo decompiled the games, built them into the project and brought AI into the app and the games.
2nd Place
Southwest Collegiate Cyber Defense Competition, 2nd Place
Baylor placed second at the Southwest regional defending a live enterprise network against a professional red team. Orojo served as network security lead (2023–2025) and now coaches the team.
$ ls talks/
Recent Presentations
Pitch2026-10-24
Kyerɛ: Watching isn't knowing
Pitch for Kyerɛ, a study app where watching a lecture fills at most half of a concept radar and the rest is earned by answering alone and by teaching Fen, an AI student who knows only what you told it. Team Barney and Friends.
Talk2026-09-26
From Measurement to Co-Adaptive Defense: Autonomous Vulnerability Remediation on Live Systems
A 15-minute research talk in the workshop's Program Analysis, Verification, and Network Defense session. It covers SysRepair-Bench (313 dual-objective scenarios across six suites), where agents lose ground (compensating controls at 44.2% success@5), NeuroPlan's reviewable plans, and red/blue self-play as the next step.
$ git log --papers
Latest Publications
Autonomous System Vulnerability Remediation: A Survey of Agentic AI, Reinforcement Learning, Benchmarks, and Operational Safety
SysRepair-Bench: A Benchmark for AI Agents' Ability to Remediate Real-World System Vulnerabilities
SysRepair: A Benchmark and Multi-Architecture Approach to Autonomous Vulnerability Remediation
Autonomous system vulnerability remediation: A survey of agentic AI, reinforcement learning, benchmarks, and operational safety
Client-Side Video Analysis for Content Moderation with MoQ
An Evaluation of ABR Switching for Time-Shifted Clients in MoQ
$ tail news.log
News
la backdoor site launches
Our new lab site is live.
$ ssh visitor@labackdoor
Poke Around in the Shell
type help to see commands, or read about the lab
la backdoor // security research lab
type `help` to get started, or `ls` to look around.
visitor@labackdoor:~$