This section is the umbrella for the platform's security architecture — the
defense-in-depth model that runs across cloud platforms (AWS, Snowflake, Databricks),
network perimeters (firewalls, fail2ban, WAF), and AI/ML inference paths (prompt-injection
defense, confidential compute, audit-log integrity). Security here is treated as a
system-design constraint, not a checklist: classification drives routing,
identity gates every call, and every action is recorded in an append-only log that can be
replayed during a compliance review.
The diagram below shows the seven layers of the model. Each layer is a control surface
with its own primitives — the cloud-specific names appear inside each layer box.
The subsections that follow link out to the per-platform deep dives and the two
cross-cutting reference pages (OWASP Top 10
for LLMs and Zero-Trust AI
Architecture) that span all platforms.
1. Defense-in-Depth Layered Model
Seven layers, top to bottom. Every inbound request crosses every layer; every outbound
response carries audit evidence back through them. Misses at one layer should be caught
at the next — that is the defense-in-depth invariant.
How to read this diagram: each layer is an enforcement point that can
stop a bad request on its own; the arrows represent the legitimate request path. A
prompt-injection attempt that bypasses the network layer should still be caught at the
inference / trust boundary. A leaked secret that bypasses identity should still be
caught by anomaly detection in the audit layer.
2. AI / ML Security
Controls specific to LLM/RAG/document-intelligence workloads — the inference and
trust-boundary layers from the diagram. See the dedicated landing page at
Security / AI & ML.
Two reference pages that span every cloud platform and apply across the entire
defense-in-depth stack.
OWASP Top 10 for LLM Applications — the canonical taxonomy of LLM-specific risks (LLM01 through LLM10), with attack scenarios, defenses, and a RAG-architecture map showing where each risk lives.
Zero-Trust AI Architecture — zero-trust tenets (verify explicitly, least privilege, assume breach) applied to AI/ML systems: identity, network, data, workload, and continuous-verification planes.
By Concern
40 pages. The primary axis is the concern, not the platform — because AWS,
Snowflake and Databricks are largely the same questions answered three times, and reading them
side by side is the point. Each entry carries its platform as a label.
Foundations (2)
Where to start: the standard threat list for LLM applications, and the architectural stance the rest of this section assumes.
The concerns that only exist once a model is in the loop — untrusted text reaching the prompt, and inference on data you cannot expose.
Prompt-Injection Defense for RAGAI/ML — In a retrieval-augmented generation (RAG) pipeline, the documents retrieved from the vector store are spliced directly into the…
Output Filtering & Canary TokensAI/ML — Ingest-time redaction ( see the redaction page ) removes PII and privileged content from prompts before inference. But an LLM can…
Confidential Computing for On-Prem InferenceAI/ML — When privileged matters are pinned to an on-prem model, the trust boundary shrinks to the physical server running the weights. But…
Homomorphic Embedding SearchAI/ML — In a conventional RAG pipeline, the user's query is embedded on the application server and then sent in plaintext (as a vector of…
Data Protection & Privacy (11)
Keeping the sensitive parts out of what the model, the analyst, or the log ever sees. The three platform answers to masking sit together here rather than in three folders.
Format-Preserving Encryption (FPE)AI/ML — AES ciphertext of an SSN looks like a 256-bit blob; a downstream system expecting cannot validate it, index it, or route on it. In…
Differential Privacy for AggregatesAI/ML — A firm running document intelligence across hundreds of matters will eventually want cross-matter analytics: "how many contracts…
k-Anonymity & l-DiversityAI/ML — Releasing a dataset — a discovery set, a training corpus, a research extract — is different from running a query against one. Once the…
Row Access and Masking PoliciesSnowflake — Row access policies and masking policies are Snowflake's data-level access controls. A masking policy rewrites the value of a…
Tag-Based Masking and ClassificationSnowflake — Attaching a masking policy to every sensitive column by hand does not scale past a few dozen tables. The tag-based approach inverts…
Row Filters & Column MasksDatabricks — Unity Catalog provides two fine-grained access controls that go beyond the table-level grant: row filters hide entire rows from a…
Amazon MacieAWS — Amazon Macie is a managed sensitive-data discovery service for Amazon S3. It uses ML and pattern matching to find PII, PHI, financial…
Identity & Access (3)
Who can do what. One page per platform — the same question, three implementations.
AWS Identity and Access Management (IAM)AWS — AWS Identity and Access Management (IAM) is the foundational service for controlling who and what can access AWS resources. It manages…
Trust Center and RBACSnowflake — Snowflake's authorization model is pure RBAC: privileges are granted to roles, roles are granted to users (or other roles), and a…
Unity Catalog SecurityDatabricks — Unity Catalog (UC) is the policy plane for every securable on a Databricks lakehouse — catalogs, schemas, tables, columns, volumes,…
Encryption & Key Management (5)
Keys, secrets, and who holds them.
AWS Key Management Service (KMS)AWS — AWS Key Management Service (KMS) is a managed service for creating and controlling cryptographic keys backed by FIPS 140-2 validated…
AWS Secrets ManagerAWS — AWS Secrets Manager stores, rotates, and retrieves secrets — database credentials, API keys, OAuth tokens — via a managed service with…
Secrets & Key RotationAI/ML — A legal-tech platform accumulates secrets quickly: per-tenant encryption keys, vault data-encryption keys, model-provider API tokens,…
Encryption Key HierarchySnowflake — Snowflake encrypts everything: micro-partition files at rest, intermediate spill files on local SSD, all in-flight data over TLS, and…
Customer-Managed Keys & EncryptionDatabricks — Databricks encrypts every byte at rest with AES-256 by default. Customer-Managed Keys (CMK) , also marketed as Bring-Your-Own-Key…
Network & Isolation (7)
Keeping traffic where it belongs — from private links down to the host firewall on this box.
VPC Security Groups & NACLsAWS — VPC Security Groups (SGs) and Network ACLs (NACLs) are the two packet-filtering primitives that AWS provides inside a VPC. They sit at…
AWS Shield & WAFAWS — AWS Shield and AWS WAF together form the edge defence for internet-facing AWS workloads. Shield protects against L3 / L4 (network and…
Network Policies and PrivateLinkSnowflake — The Snowflake account URL resolves on the public internet by default — anyone with valid credentials can authenticate from anywhere.…
Workspace Isolation & PrivateLinkDatabricks — By default a Databricks workspace exposes its UI and REST API on the public internet, and worker nodes phone home to the…
Knowing what happened, proving it later, and the obligations that come with holding the data.
AWS CloudTrailAWS — AWS CloudTrail records every API call made within an AWS account — from console clicks to SDK invocations to internal…
AWS Config InspectorAWS — AWS Config continuously records the configuration of your AWS resources, evaluates them against rules, and emits a timestamped change…
Amazon GuardDutyAWS — Amazon GuardDuty is AWS's managed threat detection service. It continuously analyses VPC Flow Logs, DNS query logs, CloudTrail…
AWS Security HubAWS — AWS Security Hub is the central security-posture service for AWS accounts. It aggregates findings from native services (GuardDuty,…
Audit Logs & ComplianceDatabricks — Every privileged action in a Databricks workspace — from a successful login, to a cluster start, to a issued in Unity Catalog — emits…
Audit Log IntegrityAI/ML — An audit log is only as good as the guarantees you can make about it under adversarial review. "We log every query" is a…
Data Residency & Sovereignty RoutingAI/ML — Sensitivity-tier routing decides which model handles a request (cloud vs. on-prem). Residency routing is a second, orthogonal axis:…
Supply-Chain Security for Model ArtifactsAI/ML — On-prem inference reduces the exposure of prompt and response data — but only if the model you are running is the model you think you…
By Platform
The same pages, grouped the way the folders are — useful when you already know which
platform you are working on.