Automating Phishing Triage With Agentic AI and Knowledge Graph Memory on Databricks
Overview
| Experience | In Person |
|---|---|
| Track | Cybersecurity |
| Industry | Healthcare & Life Sciences |
| Technologies | AI/BI, Unity Catalog, Databricks Agents |
| Skill Level | Advanced |
| DOWNLOAD SESSION SLIDES | |
Phishing attacks remain a persistent threat, exploiting human trust through sophisticated impersonation. In this session, we present a Databricks-based compound agentic system using knowledge graphs, multiple agents, and MLflow to automate triage of reported phishing emails at scale, reducing analyst workload and accelerating threat response. We walk through scalable pipelines built on Spark Structured Streaming and Delta Lake, then describe how our knowledge graph provides long-term memory, storing historical emails to enable reasoning across campaigns. We demonstrate agentic orchestration with LangGraph where specialized agents use tools to analyze headers, sender reputation, body content, URLs, and behaviors before producing triage decisions with reasoning traces. We cover evaluations, observability, including hallucination detection, metrics, and tracking with MLflow. Join us to learn practical patterns that combine multi-agent reasoning with production workflows on Databricks.
Session Speakers
Virender Dhiman
/Lead Data Scientist
CVS Health
andrew HintonA
/Distinguished Engineer
CVS Health
Full Summary
Inside CVS Health's agentic AI for phishing investigation at scale
Phishing has shifted from crude spam to adversarial, polymorphic, LLM‑authored campaigns that adapt faster than static defenses. A single missed click can trigger credential theft or ransomware, yet most user‑reported emails are benign, which overwhelms analysts. CVS Health describes Langphish, an agentic system on the Databricks Platform that treats phishing triage as an investigation with memory, structure, and auditability, not as a one‑shot classification.
FAQ
The synthesis agent outputs a recommended next step with confidence scores, such as spam likelihood, phishing likelihood, and malicious intent. Clear benign cases are recommended for closure, while suspicious or high‑risk cases are escalated with supporting context from sub‑agent findings and related campaign history.
Analysts can override any decision. Overrides are captured in audit logs within the Databricks Platform and the SOAR tool, and they feed back into the knowledge graph so future triage benefits from human judgment.
No. It operates after a user reports a message, downstream of standard email defenses. The focus is automating investigative triage of reported emails, not filtering all inbound mail.
Sub‑agents perform narrow, well‑defined checks where consistency and cost matter most, so smaller models at temperature zero work well. The synthesis step requires reasoning across all findings, which benefits from a larger model.