Identity and API
Identity follows the query through document retrieval.
Risk
Cross-user or cross-organization access.
Architecture controls
Validate sessions and tokens server-side, enforce limits and derive the tenant from a trusted identity.
THIAGO LIMA SONETI DA SILVA / CYBER ARCHITECT
Thiago Lima Soneti da Silva (Thiago Silva) — Cyber Architect and technical lead at Lima e Silva Tecnologia.
In focus · November 2026
Reliable automation:
every result needs evidence.
Starting a task does not prove completion. Design agents with idempotent actions, recoverable checkpoints and delivery confirmation, while preserving access boundaries for tools and data.
Explore my profile ↗ARCHITECTURE / AI SECURITY
Secure RAG
From question to answer, with controlled access.
Select a component to explore.
Query → identity → retrieval → model → response
Ingestion → index • Auditing spans the entire flow
Authorized retrieval
Semantic search only within content allowed for this identity.
Risk
A relevant document may still be unauthorized.
Architecture controls
Apply ACL and tenant restrictions before sending context to the model; propagate permission changes and revocations to the index.
Orchestration and model
Combines the question with retrieved passages to generate an answer.
Risk
Malicious instructions embedded in documents.
Architecture controls
Treat documents as untrusted data, separate instructions, limit context and do not grant privileged tools to the model.
Answer and sources
Returns an answer with verifiable references to the content used.
Risk
Data leakage or claims unsupported by sources.
Architecture controls
Check data exposure, escape output, link authorized sources and acknowledge missing evidence.
Ingestion and index
Approved documents are classified, split and indexed.
Risk
Poisoned content or stale permissions.
Architecture controls
Validate origin and format, scan files, record ownership and ACLs, encrypt and remove copies after revocation or deletion.
Audit and operations
Correlates access, retrieval and generation decisions.
Risk
Missing evidence or sensitive data in logs.
Architecture controls
Record decisions and correlation IDs with data minimization, defined retention, restricted access and alerts.
Illustrative reference architecture. This interaction explains security decisions; it does not run queries or send data to models.
Explore architecture decisions
- AWS and GCP: turn shared responsibility into tasks
Assign ownership for each control.
- Authentication: design beyond login
Device changes and recovery also establish trust.
- CI/CD: every stage carries a trust boundary
Building code and authorizing production are different decisions.
- RAG: revocation must reach retrievable content
Removing source access may leave usable copies.
- BOLA and IDOR: review the actual object
Knowing an identifier is not authorization evidence.
- Additional encryption: which threat does it address?
Application encryption changes exposure boundaries and operations.
Recent articles
Reliable automation: a trigger does not finish the job
Separate execution, durable results and delivery; recover individual steps and test retries without duplicating business effects.
Athena: SQL text needs its own confidentiality boundary
A query can reveal business intent without exposing its results. Classify and test query text as a separate asset.
SageMaker: failure paths need a security contract too
Failure logs can cross the boundary between code and credentials. Test diagnostics, access and response without losing evidence.
Browse by topic
Security architectureDesign decisions, trust boundaries and threat analysis.5 articlesCloud Security: AWS and GCPCloud architecture, networking, containers and security responsibilities in AWS and Google Cloud.10 articlesIdentity and accessAuthentication, authorization and the access lifecycle for users and services.11 articlesApplication securityControls for inputs, responses, uploads and web application operations.7 articlesAPI securityObject access control, integrations and operation integrity in APIs.4 articlesDevSecOps and SSDLCSecurity in development, pipelines, dependencies and vulnerability management.9 articlesData securityEncryption, isolation, privacy and data access management.5 articlesAI, LLM and RAG securityTrust boundaries, agents, prompt injection and information protection in AI systems.7 articlesSecurity operations and evidenceLogs, detection and evidence to investigate incidents and assess controls.2 articles
Subscribe via RSS
Get new articles in your RSS reader. Copy the feed address and add it to Feedly, Inoreader or another reader.
AI-assisted curation · Sources: Amazon Builders’ Library — Making retries safe with idempotent APIs · AWS Lambda — Asynchronous errors and retries
Independent projects
Turning ideas into working projects.
Meet Privacy Guardian · Code Provenance Guardian · CVE Hunter · Agentes de Segurança
Cloud Guardrail Assurance · MCP Authorization Security Lab · Security Architecture Decision Lab · Security Resilience Lab
Explore projects ↗