← All posts

Platform & Architecture

338 posts on platform & architecture.

ShieldGemma: A Production Acceptance Guide for Image Safety

ShieldGemma is Google''s open-weight family for policy-based safety classification. ShieldGemma 2 evaluates images against supplied safety policies and returns Yes or No token probabilities. Production approval should pin the model, policy, threshold, test corpus, route, failure behavior, and reviewer rather than treating a safety label as authorization.

ai-securityllm-securitypolicy-enforcementdevsecopsauditarchitecture
Read post →

A2A Agent Card Signing: Build a Verifier Pack

A2A Agent Card signatures use RFC 7515 JSON Web Signature fields over a canonicalized card representation. A safe rollout needs deterministic canonicalization, trusted-signer resolution, signature verification, tamper tests, unknown-signer behavior, rotation, unsigned-card policy, and evidence capture. This guide turns those requirements into a verifier pack for discovery and trust decisions.

ai-securityagentic-aiidentity-and-authorizationzero-trustauditarchitecture
Read post →

Microsoft Copilot Studio Security: A Deployment Acceptance Matrix

Copilot Studio security spans environments and maker access, plus connectors, agent authentication, publishing, audit events, data policies, downstream actions, and destination controls. A deployment review needs evidence that each control behaves in design time and at runtime. This guide provides an acceptance matrix for controlled publishing, connector restrictions, identity propagation, policy timing, denials, and evidence export.

ai-securityagentic-aiidentity-and-authorizationpolicy-enforcementauditarchitecture
Read post →

GPT-OSS-Safeguard Needs a Production Acceptance Pack

GPT-OSS-Safeguard is an open-weight safety reasoning model that classifies text against a policy supplied by the deployer. Production approval requires more than a model download: pin the model and Harmony format, version the policy, test a labeled corpus, define output handling, measure the selected reasoning effort, protect raw reasoning, and assign escalation owners. This guide packages those decisions into a production acceptance record.

ai-securityllm-securitypolicy-enforcementdevsecopsauditarchitecture
Read post →

Claude Code Security Needs a Deployment Acceptance Record

Claude Code security depends on the effective settings source, permission-rule precedence, sandbox scope, gateway routing, and controls around local execution. A deployment acceptance record freezes those facts for one version and surface, then proves them with denied, prompted, sandboxed, and route-pinning tests. This guide defines the record, the test fixtures, the evidence joins, and the boundary between routed model traffic and local developer activity.

ai-securityllm-securitydevsecopsidentity-and-authorizationpolicy-enforcementaudit
Read post →

Microsoft Entra Agent ID: A Deployment and Evidence Guide

Microsoft Entra Agent ID gives AI agents purpose-built identities built from blueprints, with autonomous and interactive OAuth patterns, sponsor and owner fields, inherited policy, and agent-specific logs. This guide turns that object model into a deployment worksheet and a kill-switch test, then shows where verified Entra context can govern a routed LLM request.

agentic-aiidentity-and-authorizationzero-trustai-securityarchitecture
Read post →

Copilot Connectors Security: Synced and Federated Acceptance Tests

Microsoft 365 Copilot connectors use two security models. Synced connectors ingest external content into Microsoft Graph, while federated connectors retrieve source data through MCP at query time and may expose write tools. This guide turns that distinction into an acceptance matrix for data movement, identity, permissions, audit evidence, disablement, and routed LLM traffic.

ai-securityagentic-aiidentity-and-authorizationpolicy-enforcementauditarchitecture
Read post →

NLWeb Security: A Route-by-Route Deployment Guide

NLWeb exposes conversational website data through /ask and /mcp, then joins retrieval, response generation, memory and optional account services. This guide assigns each route to its real control owner and shows where an authenticated outbound LLM call can use DeepInspect without confusing model-request policy with NLWeb application security.

ai-securityllm-securityagentic-aiidentity-and-authorizationpolicy-enforcementarchitecture
Read post →

Glean Security Review: Trust Center, Access and AI Egress

Use Glean’s Trust Center as the starting point, then test connector access, source permissions, identity propagation, and the model egress decision. A controlled answer request should leave a record that ties the originating user, source reference, model route, policy version and outcome together.

ai-securityllm-securityzero-trustpolicy-enforcementidentity-and-authorizationarchitecture
Read post →

Claude Connectors Security: Controlling Remote MCP Server Access

Claude connectors give an application a route to approved remote MCP servers. Secure deployments need a defined server inventory, authorization at the MCP server, narrow tool permissions, trusted connector configuration, and request-level policy for the LLM traffic that begins the flow. This article maps those controls to their actual boundaries.

ai-securityagentic-aillm-securityprompt-injectionidentity-and-authorizationpolicy-enforcement
Read post →

Cursor Audit Logs and AI Request Evidence

Cursor teams need evidence that connects each AI request to an authenticated actor, policy decision, and timestamp. This article separates provider administration records from independent request-layer evidence and maps the HTTP controls that regulated enterprises can verify during a security review.

ai-securityllm-securityauditinline-enforcementpolicy-enforcement
Read post →