Local ONNX Model

Sentry's Layer 1 uses a local ONNX model to classify suspicious payloads that heuristics cannot decide with confidence. Local inference via ort (ONNX Runtime), without depending on an external API.

ThreatModel trait

// sentry-ai/src/threat.rs
#[async_trait]
pub trait ThreatModel: Send + Sync {
    fn name(&self) -> &'static str;
    fn version(&self) -> &str;
    async fn classify(&self, evt: &Event) -> anyhow::Result<ThreatVerdict>;
}

The onnx feature is optional and disabled by default. Enable it via a Cargo feature at build time:

[dependencies]
sentry-ai = { version = "0.1", features = ["onnx"] }

Training pipeline

1. Dataset: catalogued malicious payloads (SQLi, XSS, RCE, log4shell)
   • benign traffic.
2. Offline training (Python, in tools/train) — produces a versioned ONNX model.
3. Versioning: models in models/ (e.g. models/sentry-payload-v1.onnx).
4. Deploy: binary loads the model at startup; hot swap via sentry model reload.
5. Feedback loop: incidents confirmed by the decider become retraining dataset.

Configuration

[analysis.ai]
onnx_model = "models/sentry-payload-v1.onnx"
llm_only_above = 30

CLI

sentry model status             # model version, accuracy
sentry model reload             # reloads model from disk (hot swap)

Status

• ThreatModel trait in sentry-ai/src/threat.rs.
• onnx feature exists in Cargo.toml.
• ONNX implementation pending (F2.1/F2.2) — no tools/ yet.