Security Operations AI Skill

Turn your coding agent into a security engineer

One knowledge file that builds threat detection, hunting and response inside Snowflake.

Kevin Keller · kevinkeller.org

Why it matters

Real threats slip through while analysts triage noise

  • SOC analysts triage hundreds of alerts a day, most of them false positives.
  • Most need the same manual steps: query, correlate, check, escalate or dismiss.
  • Credential stuffing at 2 AM or slow exfiltration over weeks goes unnoticed.

No human keeps vigilance at machine scale.

The idea

Give the agent the security knowledge, then ask for what you need

A file of about 2,800 lines encodes Snowflake's security views, detection patterns and three major frameworks.

Agent-agnostic

One skill file works in any coding agent

  • Built for Snowflake Cortex Code.
  • Works the same in Claude Code, Cursor, Continue.dev and Windsurf.
  • Any agent that reads context files can use SKILL.md.

No lock-in to a single agent or vendor.

Three frameworks built in

Every detection maps to a framework your auditors know

FrameworkCoverage
OWASP Top 10:2025All ten categories, with SQL detection patterns
MITRE ATT&CKTactics mapped to Snowflake views, from Initial Access to Impact
NIST CSF 2.0Govern, Identify, Protect, Detect, Respond, Recover

Five detections

Ready-made patterns for the attacks that matter most

  • Brute force: repeated failures on one account
  • Credential stuffing: many usernames from one source
  • Impossible travel: logins moving faster than possible
  • Data exfiltration: volumes far above a user's baseline
  • Privilege escalation: unusual grants, such as ACCOUNTADMIN

The pipeline

Events become alerts in 1 to 5 minutes, with no orchestration code

  1. Ingest via Snowpipe or REST API.
  2. Raw events in a Dynamic Table, 1-minute lag.
  3. Normalise and enrich, 2-minute lag.
  4. Score and alert, 5-minute lag.

Snowflake handles refresh, ordering and incremental processing.

Where alerts go

Analysts, agents, dashboards and your existing SIEM

  • Interactive Tables for analysts to filter and acknowledge.
  • Cortex agents for natural-language questions.
  • Streamlit dashboards for live monitoring.
  • Webhooks to Splunk, Sentinel, CrowdStrike, ServiceNow or PagerDuty.

Talk to it

Ask security questions in plain language

"Which users don't have MFA enabled and have ACCOUNTADMIN access?"
  • Threat hunting, compliance reports and incident timelines on request.
  • The agent generates the SQL or Python. You review and run it.

Try it

15 min

to deploy the full demo

Five scripts create the schema, generate 10,000+ events with injected attacks, build the pipeline, analyst tables and a Streamlit dashboard.

The takeaway

Your agent already writes code. Now it knows security.

Read the full article

Kevin Keller · Open-source skill on GitHub