Snowflake XDR
A security operations centre that runs itself
Cortex AI does the routine SOC work inside one Snowflake account. Humans approve what could hurt.
Kevin Keller · kevinkeller.org
Why it matters
The bottleneck is people, not compute
- SIEM pricing is per ingest, so teams decide what not to log
- Detection rules are slow to write and break silently when log formats change
- Noisy rules get switched off and analysts drown in triage
- An AI that dismisses alerts can also learn to hide an attack
The approach
Bring the security platform to the data
- Logs arrive over standard OpenTelemetry into one table
- A mapping step turns every source into Splunk CIM-shaped tables
- Every detection is SQL over those tables, so one rule works on any EDR
- Cortex AI writes rules, mappings, triage opinions, notes and reports
From log line to closed case
Nine steps, one human click
- A new source connects; Cortex writes its log mapping
- Rule packs install in shadow; canaries prove they fire
- The Detection Autopilot writes rules for the top coverage gaps
- Findings get confidence scores, then group into incidents
- P1 and P2 cases investigate themselves; containment waits for one human click
By the numbers
Built from standard data-platform parts
| Rule packs | 119 rules in 12 packs |
|---|---|
| Sigma | 1,089 community rules, all 3,802 branches verified to fire |
| Fast lane | 12–22 seconds from EDR alert to finding |
| Ask XDR agent | 27 tools, 25 of them custom actions |
| Fresh build | about 18 minutes, one command |
The Detection Autopilot
Autonomy is a setting per type of change
| Change type | Default |
|---|---|
new_rule | Shadow 24 h, then promote |
tune_threshold | Live immediately |
disable_noisy | Propose only |
new_parser | Propose only |
It never deletes data or rules, and every change rolls back in one click.
The key idea
False-positive reduction that can't silently hide attacks
Only human verdicts train the suppression model. A dismissed finding that another rule corroborates re-opens itself.
Guardrails in SQL
Containment refuses unless every check passes
- A human must have approved the run
- The target must be on the response allowlist
- It cannot contain the calling user
- Admins are never contained automatically
- Every action records what it changed, so undo restores it
Mission Control
The human's job shifts from doing to supervising
- Needs you: one ranked queue of approvals and spot checks
- What the AI did: a ledger with an undo button on every change
- Kill switch: anyone on shift can pause all automation
- Earned autonomy: canaries and track records decide how much rope the AI gets
Before anyone uses it for real
The gaps worth naming
- Prompt injection through attacker-controlled log fields
- The SQL guard is a regex: a seatbelt, not a sandbox
- Approver role and four-eyes checks belong in the procedure
- Live connectors, cost and real data volumes are untested
Takeaway
Fully automated by default, human-gated where a mistake is expensive, one button to stop it all.
Personal project, not an official Snowflake product, not endorsed by Snowflake, not meant for production