Hypervisor-zero is an independent AI safety and security research project focused on making advanced AI agents safer, more controllable, and more reliable in real-world environments.

The project investigates how AI agents can be monitored, constrained, and contained while carrying out tasks, particularly when they encounter malicious instructions, adversarial attacks, or situations that could lead to unsafe actions.

this research will combines AI agent runtime security, swarm-agent cyber defense, adversarial testing, and independent AI control evaluations. I'm developing ControlBench, a benchmark designed to measure whether AI agents actually respect security policies, whether their actions are successfully blocked when necessary, and whether protected systems and data remain safe

The project is led by an independent researcher and third-year software engineering student, with a background in AI research, dataset development, and model fine-tuning.

Current stage: Hypervisor-Zero has a working prototype, a researcher dashboard, and an evolving technical architecture. We are currently improving our evaluation pipeline, building and validating research datasets, developing adversarial test scenarios, and preparing controlled experiments. The next stage is to establish reliable baseline results and conduct more extensive evaluations.

Request access at:

Hvisorzero@gmail.com

Or

info@hypervisorzero.xyz