Closed early
Tarnhaven
A study of whether defenses built against automated, model-driven attackers actually stop autonomous AI agents. It opened on 2 October 2026 and I closed it early, because nobody took part.
The question
Tarnhaven was a synthetic company site with no real data, built so that volunteers could point their own AI agents at it. The question was simple: do defenses designed for automated, model-driven attackers hold up against real autonomous agents, or only against the attackers we imagined?
What happened
The window was planned for 28 days. Nobody took part. With zero participants there is nothing to analyse, and keeping idle infrastructure running would only have burned free cloud credits that I would rather keep for the next experiment. So I closed it early.
The honest reading is that the defenses were never tested at all, which means the study says nothing about them, good or bad. A testbed nobody knows about is not a study. Whether I run a second round is an open question.
What is left
You can find the architecture write-up here: AI Defense Study: how the two machines were separated, what was meant to be measured and why. The study setup and outcome are documented above; no live participant data was collected.