Planned, not yet built. The first of two projects I am working towards as part of learning AI security properly, rather than reading about it.

The approach I want to test: measure normalised facial landmark jitter with OpenCV and MediaPipe Face Mesh across a baseline window, then compare it against a physics-based challenge-response, instead of training a CNN on a spoof dataset. A classifier learns the spoofs it was shown and fails on the ones it wasn’t. The question worth answering is whether a photo or a replay can reproduce the micro-motion signature of a live face reacting to an instruction it could not have anticipated.

Planned, not yet built. The second project in the same AI security track, and the one I expect to learn more from.

The intent: a harness that runs structured prompt injection suites against a model and reports pass and fail rates per injection category, tracks how those rates move across model versions, and records where guardrails break down under adversarial input. Vendors publish claims about robustness; the point of building this is to be able to check them.

A security scanning tool written in Python that enumerates cloud infrastructure and checks for common misconfiguration patterns. Surfaces exposure before an adversary can exploit it.

A segmented security lab built on a Proxmox hypervisor, with the attacker workstation on separate hardware from the monitored network. VLANs and an inline IDS isolate attacker, victim and management traffic. Sysmon and Windows Event Forwarding feed a Wazuh SIEM, with every detection mapped to MITRE ATT&CK.