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.

Models

Galvanize 2 (current): multilingual prompt injection classifier, about 141M parameters, Apache-2.0. Fine-tuned from Horizon-Labs/prompt-injection-guard-small (Apache-2.0). On a private held-out test set of 3,309 examples, at the serving threshold of 0.662: 83.4% recall and 0.81% false positive rate.

Model: usezn/Galvanize-2 · Launch notes: Introducing Galvanize 2

Galvanize-60M (previous): 60M parameter prompt injection classifier distilled from ModernBERT base, Apache-2.0. It remains available and serves as the automatic fallback.

Model: usezn/Galvanize-60M · Launch notes: Introducing Galvanize-60M

Dataset

zn Prompt-Injection Evaluation Corpus: binary text classification corpus of prompt injection attacks and hard benign negatives, with per-row source, license and language. CC-BY-4.0.

Dataset: usezn/zn-prompt-injection-bench

Links

• Website: usezn.com
• Code: github.com/usezn/zn, the open-source local security layer for AI coding agents
• Security reports: security@usezn.com