A few pieces I wanted to show people, mostly on software security. Some of it is research, some is coursework from my degree, and the rest comes from my own projects.
A framework (SAIF) for folding AI-accelerated development into an agile workflow without losing security, architecture, or reliability. Grounded in a security experiment across four AI coding models and a semester building inside the ToSee AS startup. A group thesis with Bendik Hagen and Krister Eriksen.
How an adversary actually takes down a power grid: the threat landscape from Industroyer-class ICS malware to supply-chain compromise and social engineering against operators. A look at what defense-in-depth for critical infrastructure really requires once the attacker is already inside. Built from my cyberwarfare coursework.
A full network and server designed and secured from scratch: the topology, IP scheme, firewall and ACL rules, the server, and backups, all worked out, justified, and costed.
How a single secret bug gets priced, who pays, and why the number keeps climbing. The mechanics of the zero-day market, from white-market bounties to seven-figure broker deals, the iOS-versus-Android price flips, and the state sitting as buyer, referee, and regulator all at once.
The arms race over AI-generated text: watermarking, post-hoc detection, and the cheap paraphrase that breaks both. Backed by a small local experiment, two detectors collapsing from ~0.95 to ~0.65 AUROC after a single rewrite, and the case for provenance over detection.
We are getting good at detecting when a frontier model is dangerous, but not at stopping it from shipping. The containment and proliferation problem: once a capable model exists, the weights leak and copies spread, and the real question is whether governance or international treaties can hold the line.
What the actual studies, benchmarks, and enterprise telemetry say about the security of AI-generated code, separating the hype from the reality. Where the real risk concentrates, which vulnerability classes recur, and why the model is rarely the whole story.