Hermes Katana
A defense-in-depth security toolkit for LLM agents that tracks where every piece of text came from, scans it for prompt injection, applies policy before a tool runs, and scrubs outbound secrets.
Hermes Katana attaches a source, such as user input or fetched web content, to every character of text using a taint-tracking model inspired by Google DeepMind's CaMeL paper, so that when strings from different sources are combined, sliced or transformed, each resulting character still carries its origin and a tool call built from tainted, dangerous content can be denied automatically. Runtime policies, written in YAML, decide what happens for clean, tainted, dangerous or unknown-origin tool calls, including routing some calls to a human for approval.
Injection detection uses a small distilled MiniLM classifier by default, with a larger DeBERTa-v3-large model available for higher accuracy, and a Proving Ground harness runs empirical, multi-model attack-effectiveness tests with explicit gates against false positives and against regressions in known attacks. The base install is intentionally small and works without downloading any model, with larger model and dataset artifacts kept on Hugging Face rather than in the repository.