Research.

Day Zero is the research arm of Efficient Frontier Labs: a field-report series on where AI agents pressure-test the infrastructure they run on, and the primitives that turn out to matter. Each report begins as an authorized adversarial security experiment — live systems, real attackers, machine-speed exploration — and ends as a finding anyone can check.

The thesis is plain. Machine-speed systems don't need more dashboards; they need proof. We run the experiments in the open, in Oakland, and publish what broke and what held — counts, telemetry, and the disagreements included.

• № 01AuthorizedEvery experiment is bounded and consented.
• № 02AdversarialReal attackers, autonomous and human.
• № 03InstrumentedGround-truth telemetry, not dashboards.
• № 04On the recordDated revisions; contested findings kept.

Each experiment is authorized, bounded, and run live. Red teams — human operators and the autonomous systems they direct — work against real deployments while a judge aggregator polls every team's event journal on a fixed cadence and persists the results. Blue teams defend, redeploy, and respond. What we publish is the trace.

We treat observability as part of the defense, not a layer beside it. Where application-plane scores depend on the substrate underneath, we say so; where a finding is contested, we keep the disagreement in the record and resolve it in dated revisions. The reports are written to be reproduced from the cited bundles, not taken on faith.

The series runs out of 1900 Broadway in Oakland, alongside the weekly Office Hours session. New volumes are announced there first.