KOR IT LAB
Build. Measure. Break. Improve.
The KOR IT LAB provides controlled environments for testing emerging security, data and AI architectures before they become real-world systems. It exists so that what we recommend has been run, measured and broken somewhere first.
Domains
Four laboratories, one environment.
Each domain asks different questions of the same host, which is why their findings connect.
AI Lab
Local inference, open models, agent architectures, memory and the performance characteristics of running all of it on owned hardware.
- Local inference
- Open models
- Agent architectures
- Memory
- +2
Security Lab
Agent security, memory provenance, runtime policy, tool boundaries and what any of it looks like from a SOC.
- Agent security
- Memory security
- Runtime policy
- Tool security
- +2
Data Lab
Security data architecture, governance, metadata, lineage and data quality — modelled and tested before it reaches an estate.
- Data architecture
- Security data
- Governance
- Metadata
- +2
Observability Lab
Infrastructure, AI, agent, network and security telemetry — designed together, because the questions cross all five.
- Infrastructure telemetry
- AI telemetry
- Agent telemetry
- Network telemetry
- +1
The environment
One host. Owned, instrumented, and stated plainly.
One host, operated by KOR IT, used for every experiment below. Measurements taken here characterise this machine and nothing beyond it.
Hardware
- CPU
- AMD EPYC 7451
- Cores / threads
- 24C / 48T
- Memory
- 256 GB ECC
- Storage
- NVMe
- NUMA domains
- 4
- Virtualisation
- Proxmox
Stack
- Inference
- llama.cpp · Qwen-family open models · local embeddings
- State
- PostgreSQL · Qdrant · MinIO
- Observability
- Prometheus · Grafana · Loki · Tempo
Why a lab at all
Most of what is written about agentic AI security is written by people who have not had to run one. The LAB exists so that KOR IT's recommendations come from a system that was actually built, instrumented and pushed until something failed — rather than from a threat model that has never met an implementation.
That has a cost we state openly: a single host, a small number of operators, and no external adversary. Untrusted content is introduced by us, which means the threat model is exercised against attacks we already thought of. Every lab page carries its own limitations section for exactly this reason.