Appendix

Glossary, scope and assumptions, and references.

Glossary

Term Meaning in this report
ACCOUNT_USAGE The schema in the SNOWFLAKE shared database that exposes account-level audit and inventory views (LOGIN_HISTORY, QUERY_HISTORY, STAGES, SHARES, etc.). Latency is approximately 45 minutes.
Snowflake Trail Newer near-real-time audit surface; lower latency than ACCOUNT_USAGE and covers some events not visible there. The right source for time-sensitive alerting pipelines.
Trust Center Snowflake's first-party security-posture scanner. The "Security Essentials" pack flags missing MFA, missing network policies, and over-privileged service users.
Horizon Catalog Snowflake's governance plane: data classification and tagging, column- and row-level security, masking and tokenization policies, and Cortex AI Guardrails.
Cortex Code Agentic CLI that operates against a developer's local filesystem, can run shell commands, and holds cached Snowflake tokens. Subject of CVE-2026-6442.
Cortex Analyst Text-to-SQL over a curated semantic model. Constrained to SELECT at the Analyst boundary; an Agent that wraps Analyst is not so constrained.
Cortex Search Embedding-based search over customer-indexed documents. Indexed content becomes a potential indirect-prompt-injection delivery channel for downstream Agents.
Cortex Agents Orchestration layer that combines Cortex Analyst, Cortex Search, and tool calls (including MCP). Inherits the trust profile of every constituent component.
NAAAPS Native App Anti-Abuse Pipeline Service — Snowflake's automated review pipeline for every new Native App version or patch. Auto-approve or escalate to manual review.
SPCS Snowpark Container Services — container hosting inside a Snowflake account, with strict default network isolation modulated by customer-managed EXTERNAL ACCESS INTEGRATIONs.
Storage Integration Snowflake object that binds the warehouse to an IAM role (or Azure managed identity / GCP service account) in the customer's cloud account, enabling external-stage reads and writes.
External Function SQL-callable function whose implementation lives in AWS API Gateway, Azure API Management, or Google Cloud Run. The execution role of the backing function is the trust boundary.
Direct Share Cross-account read access granted by CREATE SHARE and ALTER SHARE ... ADD ACCOUNTS. Data motion runs server-side; no QUERY_HISTORY entry on the consumer side for the data itself.
Replication Group Snowflake primitive for bulk replication of databases and integration objects across accounts and regions. Useful for DR; abusable for query-history-bypass exfil.
Reader Account Provider-controlled Snowflake tenant created for a consumer who does not have their own Snowflake contract. Misconfigured reader accounts can serve as unmanaged exfil destinations.
PAT Programmatic Access Token — long-lived bearer token scoped to a specific role and (optionally) a specific network policy. GA 2024/2025.
SCIM System for Cross-domain Identity Management — the IdP-side provisioning channel that syncs users and group memberships into Snowflake. The SCIM bearer token is the trust article.
MCP Model Context Protocol — open protocol for exposing tools to LLM-driven agents. Cortex Agents can consume MCP tools; tool descriptors and tool output flow into the Agent's grounding context.
FROSTBITE The reconnaissance utility Mandiant attributed to UNC5537 during the 2024 Snowflake breach campaign — enumerated users, roles, current IP, session IDs, and organization names via SQL.

Scope and threat-model assumptions

In scope

  • Snowflake authentication and identity (password, MFA, key-pair, OAuth, SAML, SCIM, PATs).
  • Cortex AI surface — Code, Analyst, Search, Agents, Document AI — including MCP tool integrations.
  • Native Apps and the Marketplace consumer / provider trust boundary.
  • Snowpark Container Services and Streamlit-in-Snowflake.
  • External Functions and Storage Integrations as outbound trust paths into customer cloud accounts.
  • Data Sharing primitives: Direct Shares, Replication Groups, Reader Accounts.
  • Execution primitives: stored procedures, UDFs, Tasks.
  • Detection surface: ACCOUNT_USAGE, Snowflake Trail, Trust Center, Cortex AI audit events.

Out of scope and known limits

  • Server-side Snowflake service vulnerabilities. Multi-tenant SaaS issues are remediated server-side and rarely receive CVEs. The Trust Center and platform security bulletins are the authoritative signal for service-side posture.
  • Cortex Guardrails detection-quality characterization. Guardrails is referenced as a detection layer; this assessment does not empirically measure its false-positive / false-negative rate on a corpus of public indirect-prompt-injection payloads.
  • SPCS egress-filter depth. Service-spec misconfiguration is the modeled threat (Chain H); this report does not characterize whether the egress inspection layer is DNS-only, SNI, or full L7.
  • Snowflake Trail vs. ACCOUNT_USAGE event-coverage diff. Both audit surfaces are referenced; a precise field-by-field mapping is a follow-on effort.
  • Vendor-side disclosures about Cortex inference egress. Exact payload shape, retention, and cross-invocation caching at Anthropic / Azure-OpenAI are vendor-side facts to obtain through a DPA review, not red-team activity.

Threat-model assumptions

  • Attacker is financially motivated, capable, and patient — the UNC5537 profile, augmented with 2026-era AI-aware tradecraft.
  • Initial-access dictionary includes infostealer logs, AiTM-captured federated cookies, IdP compromise, CI/CD compromise, and developer-host compromise.
  • Customer has a SIEM that receives connector logs and at least some Snowflake audit events.
  • The customer's IdP (Entra / Okta / Ping) is trusted for human users; the IdP is itself a soft target for chain D.
  • Snowflake's platform-default controls as of 2026 are in place but customer adoption of opt-in hardening (network policies on key-pair users, MFA-verified federation, audit replication across regions) is uneven.

Companion material in this repository

  • docs/analysis/snowflake-platform-attack-surface-2026.md — the analytical companion to this report.
  • docs/analysis/entra-2026-state-of-play.md — relevant for chain D (federated-IdP compromise).
  • docs/analysis/aitm-kit-market-2026.md — relevant for chain A and chain D credential-theft initial access.
  • docs/methodology/llm-attack-modeling.md — Cortex chains B and I directly extend this methodology.
  • docs/methodology/ci-cd-attack-modeling.md — relevant for chain C (provider-account compromise) and chain F (CI-runner compromise).