Armorer Guard Learning Loop: local live feedback without model drift
We just pushed the Armorer Guard Learning Loop.
It is a Rust-native hybrid live learning layer for local agent security enforcement:
- feedback adapts local enforcement immediately
- global model improvements still go through reviewed, versioned retraining
- no scanner network calls
- no silent cloud upload
- no poisoning-by-default
New CLI modes:
armorer-guard feedback-record
armorer-guard feedback-stats
armorer-guard feedback-export --reviewed-only
inspect / inspect-json now also include scan_id, model_version, and learning_version, so feedback can be tied back to a stable local scan without requiring raw text storage.
The local overlay can suppress eligible semantic false positives after feedback, or add local block/review matches. It deliberately cannot suppress credential or dangerous policy reasons like detected:credential, policy:credential_disclosure, or policy:dangerous_tool_call.
Repo/docs: https://github.com/ArmorerLabs/Armorer-Guard
Demo Space: https://huggingface.co/spaces/armorer-labs/armorer-guard-demo
Would love feedback from people building agents or eval harnesses: where would you put this loop: prompt ingress, retrieval ingress, model output, tool-call args, or all of them?
Quick release update: the Rust CLI is now live on crates.io as armorer-guard v0.2.2.
Install:
cargo install armorer-guard --locked
This release includes the local Learning Loop (feedback-record, feedback-export, feedback-stats) so teams can adapt enforcement from reviewed local feedback without uploading prompts or silently mutating model weights.
Crate: https://crates.io/crates/armorer-guard/0.2.2
Repo: https://github.com/ArmorerLabs/Armorer-Guard