zn Research

Independent research lab for trustworthy autonomous agents. We study how guards fail inside agent tool loops and publish open weights, methods and evaluations to fix it.

Research areas

Robust detection under distribution shift, efficient encoders on CPU budgets, representation learning for structured payloads, calibration of operating points, and evaluation methodology for agent security.

Artifacts

Galvanize-60M: 60M parameter prompt injection classifier distilled from ModernBERT base, Apache 2.0. Certified 1.00% tool false positives, 91.60% out of distribution injection recall, 11.52 ms p50 CPU, native 8,192 token context, INT8 build at 176 MB.

Model: usezn/Galvanize-60M · Methods and benchmark: Introducing Galvanize-60M · Lab page: usezn.com/galvanize

Reproduce

Weights, tokenizer and quantized graph ship in the repo. The launch notes document the 32 core protocol, tool holdouts and long context suites so every number can be remeasured.