Detection Layers

Sentry combines three detection layers in a cascade: deterministic heuristics (fast, zero known false positives), a local ONNX model (for the unknown), and an optional on-demand LLM (for complex payloads without a clear verdict).

Layer 0 — Heuristics

Always runs, ~µs. Regex for SQLi/XSS/path traversal, ASN allowlist, local IP reputation. 9 detectors in sentry-core/src/heuristics.rs:

• SQLi
• XSS
• PathTraversal
• LFI
• Log4Shell
• CmdInjection
• SensitivePath
• BadCrawler
• EmptyUserAgent

Heuristics run on the URL-decoded form of the path/query (heuristics::http_text), so that encoded payloads (%27 = ', + or %20 = space) cannot bypass them.

Layer 1 — Local ONNX model

Classifier trained on malicious payloads (SQLi, XSS, RCE, log4shell). Offline training, versioned model in models/. Local inference without depending on an external API.

Layer 2 — On-demand LLM (optional, high cost)

Only for Medium events without a clear verdict; short prompt with truncated path+headers+payload. Structured response via JSON schema. Provider-agnostic via the LlmProvider trait.

Retraining

Offline pipeline consumes confirmed incidents → new model → sentry model reload.