
Finds the detection rules in your SIEM that are running blind
Open-source SIEM detection health.
Find enabled detections that are blind because their telemetry is missing, stale, late, or
schema-incompatible.
Runs locally · Read-only · No agent · No telemetry upload
Read the technical write-up · Featured in Detection Engineering Weekly · Featured in tl;dr sec #341
Real scan of a disposable Elastic lab with deliberately missing, stale, late, and unused telemetry. Open the image for the short replay, or reproduce it with make record-scan-lab.
A rule can be enabled, scheduled, and error-free after the data it needs has disappeared. deadair reads the live rule inventory, resolves each rule's inputs using the backend's native semantics, and checks the concrete sources behind them.
It catches:
deadair supports Elastic Security, OpenSearch Security Analytics, and Microsoft Sentinel.
Download a binary for macOS, Linux, or Windows from GitHub Releases, or install with Go:
go install github.com/alephnull-sh/deadair/cmd/deadair@latest
Print the read-only setup for your SIEM:
deadair setup elastic # Elastic Security
deadair setup opensearch # OpenSearch Security Analytics
deadair setup sentinel # Microsoft Sentinel
Run one setup, then verify and scan:
deadair check # verify the credential can scan
deadair scan # assess live rules and telemetry
Exit codes are stable: 0 passes the configured gate, 1 means gated findings, and 2 means the scan failed.
| Stage | What deadair does |
|---|---|
| Inventory | reads enabled detections and the inputs they declare |
| Resolve | uses native index resolution on Elastic and OpenSearch; on Sentinel, combines KQL analysis with table, watchlist, saved-function, ASIM, and mapped cross-workspace evidence |
Sentinel follows the same rule-to-source model. Its adapter also understands literal watchlists, saved functions, ASIM parsers, mapped workspaces, and summary-table lineage. When Azure provides enough evidence, deadair can show that one filtered slice of a shared table has gone quiet or that a summary pipeline has fallen behind. Those two checks are advisory; they do not change the gate. The usage guide describes the evidence rules, and the validation record records the live test coverage.
Live scan of a disposable Sentinel lab seeded with missing, stale, late, and incompatible telemetry. Open the image for the short replay. See the separate Azure conformance record for the read-only and write-denial tests.
deadair checks whether a detection's telemetry is present and healthy. It does not validate rule logic or prove that a simulated attack will fire an alert. Use static rule validation and end-to-end detection tests for those jobs.
Every verdict is limited to what the configured credential can see. JSON reports include the configured expressions, resolved sources, resolution method, assessment status, backend metadata, and capability evidence. See the usage guide for worked examples and triage.
Elastic:
export DEADAIR_ES_URL=https://es.example.internal:9200
export DEADAIR_KIBANA_URL=https://kibana.example.internal:5601
export DEADAIR_API_KEY=<read-only-api-key>
deadair check
deadair scan --json-out report.json --html-out report.html
OpenSearch:
export DEADAIR_BACKEND=opensearch
export DEADAIR_OPENSEARCH_URL=https://opensearch.example.internal:9200
export DEADAIR_OPENSEARCH_USERNAME=deadair
export DEADAIR_OPENSEARCH_PASSWORD=<password>
deadair check
deadair scan
Microsoft Sentinel:
az login --tenant <tenant-id>
export DEADAIR_BACKEND=sentinel
export DEADAIR_AZURE_SUBSCRIPTION_ID=<subscription-id>
export DEADAIR_AZURE_RESOURCE_GROUP=<resource-group>
export DEADAIR_SENTINEL_WORKSPACE=<workspace-resource-name>
# Optional: JSON allowlist for literal workspace() targets.
# export DEADAIR_SENTINEL_REMOTES=/restricted/path/sentinel-remotes.json
deadair check
deadair scan
Before deadair assesses a rule's mapped remote workspace, that workspace must have Sentinel deployed. Same-subscription mappings can prove source availability. Cross-subscription rules need runtime evidence tied to the exact rule identity. See the Sentinel usage details for the evidence rules, workspace and region limits, and Microsoft's performance guidance.
Use the documented read-only roles for Elastic, OpenSearch, or Microsoft Sentinel.
# Gate a candidate rule against live source availability.
deadair scan --rule new-rule.json
# Fail only on new regressions between reports.
deadair diff yesterday.json today.json
# Scan multiple SIEM instances from one process.
deadair scan --fleet fleet.json
# Export cached scan results as Prometheus metrics.
deadair serve --interval 5m
scan --rule isolates a backend-native candidate rule or detector from unrelated backlog. diff
works with redacted reports created with the same caller-held key. Fleet configuration references
secrets through environment variables rather than storing secret values.
The official GitHub Action wraps single-instance candidate gates for Elastic, OpenSearch, and Sentinel. It writes a job summary, uploads a redacted JSON report, and can apply a deadair policy without installing a rule. Sentinel workflows authenticate the runner to Azure first; the Action defines no Azure credential inputs.
See CI gate behavior, fleet and MSSP deployment, and the Prometheus examples for configurations to test in your own environment.
| Backend | Live validation |
|---|---|
| Elastic Security | trusted CI on 8.19.19 and 9.4.4 |
| OpenSearch Security Analytics | trusted CI on 2.19.6 and 3.7.0 |
| Microsoft Sentinel | recorded opt-in conformance in disposable UK South workspaces; see validation status |
The Sentinel conformance run is manual, not scheduled CI.
0600 on POSIX systems.--redact replaces tenant, rule, source, pattern, field, dependency, lineage, provenance,
workspace, watchlist, template, and package identifiers with keyed HMAC pseudonyms. Validated
dependency probe expressions and their KQL arguments are never serialized. A
--redact-key-file generated from random bytes also enables redaction and keeps names stable
across separate runs.Treat reports as sensitive SOC artifacts: they identify blind detections, source names, schema gaps, and unused collection.
Bug reports, sanitized fixtures, correctness cases, docs, and backend proposals are welcome. Start with CONTRIBUTING.md and use the backend RFC template for adapter work.
Apache-2.0.
| Measure | checks source freshness and timing, plus schema and storage where the backend supports them |
| Report | emits terminal, JSON, HTML, fleet rollups, and Prometheus metrics with the evidence behind each verdict |
| Finding | Meaning | First check |
|---|
| no matching source | none of the rule's inputs resolve to a visible index, data stream, or Sentinel table | pattern changes, missing integrations, and credential scope |
| all sources stale or empty | every resolved source is unusable right now | source cadence and the ingest path |
| missing fields | an Elastic rule-declared field is absent or non-searchable in one or more resolved sources after every source mapping was read | parser, package, and mapping changes |
| lag blind window | paired-event p95 ingest lag exceeds the rule's lookback margin | rule interval, lookback, timestamp override, and pipeline delay |
| partial input coverage | the complete expression resolves, but one positive selector within it resolves empty | migrations, fallback selectors, and expected alternatives; informational unless policy gates it |
| source plan incompatible | a Sentinel rule depends on a Basic or Auxiliary table that is not eligible for the analytics-rule evidence path | table plan and rule type |
| source degradation | a source is stale, empty, low-volume, or schema-drifted | source history and expected maintenance |
| unused telemetry | on Elastic or OpenSearch, data is being stored but no enabled local detection resolves to it | disabled rules and intentional collection |