← All posts

Platform & Architecture

338 posts on platform & architecture.

Azure Prompt Shields: REST Detection and Policy Boundaries

Azure Prompt Shields detects user prompt attacks and instructions hidden in documents. The standalone Azure AI Content Safety API returns Boolean attack results, leaving the application to decide how to handle each input. This guide maps the REST request, response handling, Foundry configuration, regression tests, and boundary between attack detection and identity-aware request policy.

prompt-injectionllm-securityai-securitypolicy-enforcementarchitecturecloud-security
Read post →

RFC 8693 Agent Delegation: Token Exchange at the LLM Request Boundary

RFC 8693 defines an OAuth 2.0 token exchange request that can preserve a subject, identify an acting agent, and target a specific resource or audience. For authenticated agent-to-LLM HTTP traffic, the useful pattern exchanges an upstream credential for a short-lived token that an AI gateway can verify before forwarding the model request. The RFC supplies delegation evidence, while deployment policy still owns prompt authorization and model behavior.

agentic-aiidentity-and-authorizationai-securityzero-trustpolicy-enforcementarchitecture
Read post →

Gateway API Inference Extension: Production Routing Guide

Gateway API Inference Extension adds model-aware endpoint selection to Kubernetes Gateway API. Its stable InferencePool resource groups compatible model-serving Pods, while an Endpoint Picker can choose a replica using queue state, KV-cache utilization, and adapter availability. This guide covers the HTTPRoute attachment, failure behavior, production acceptance tests, and the security boundary between inference scheduling and request policy.

kubernetesarchitectureai-securityllm-securitypolicy-enforcement
Read post →

MCP Elicitation Security: Separate User Input from Credentials

MCP elicitation lets a server request more information during an interaction. The 2025-11-25 specification separates form mode for reviewable structured input from URL mode for sensitive interactions outside the client. This guide maps identity binding, consent, URL handling, completion notifications, rate limits, token custody, and audit evidence to an implementation test plan.

agentic-aiidentity-and-authorizationai-securityllm-securitypolicy-enforcement
Read post →

Spotlighting Prompt Injection Defense: Mark Untrusted Input Before Inference

Spotlighting marks untrusted documents so an LLM can distinguish retrieved data from application instructions. Microsoft Research tested marking techniques alongside encoding against indirect prompt injection. This guide explains each technique, where to apply it in a retrieval pipeline, how to test it, and why action authorization still belongs outside the model.

prompt-injectionllm-securityai-securityinline-enforcementarchitecture
Read post →

Perplexity Enterprise Security: SSO, SCIM, and Connector Controls

Perplexity Enterprise security covers SSO, SCIM, sharing, files, connectors, and retention. Review those workspace controls, then test each routed AI request for its authenticated principal, source context, selected model destination, policy version, and decision record.

ai-securitycloud-securityidentity-and-authorizationzero-trustpolicy-enforcementdata-loss-prevention
Read post →

MCP Dynamic Client Registration: Choose the Registration Model

MCP authorization supports preregistration and RFC 7591 Dynamic Client Registration, plus Client ID Metadata Documents. This guide compares the three approaches through redirect-URI control, metadata trust, registration-token handling, client lifecycle, and fail-closed tests while keeping MCP authorization outside the LLM gateway boundary.

mcp-securityoauthidentity-and-authorizationai-agentspolicy-enforcement
Read post →

Model Cards as Security Route-Admission Inputs

Provider model cards describe intended use and limitations, along with evaluation conditions and model versions, but they are supplier evidence rather than operational proof. This guide converts model-card claims into route-admission conditions and compensating controls, with review expiry, runtime exceptions, and a versioned approval record.

model-governanceai-securityvendor-riskpolicy-enforcementaudit
Read post →

HarmBench: Build a Versioned Release-Acceptance Pack

HarmBench provides a standardized framework for evaluating harmful-behavior elicitation and consistent refusal. This guide turns a run into a versioned release-acceptance pack with behavior subsets, classifier configuration, test identity, model endpoint, policy bundle, results, limitations, and retained evidence.

ai-securityred-teamingmodel-evaluationjailbreaksllm-security
Read post →

Rich Authorization Requests for AI Agents: Carry the Grant into Policy

RFC 9396 lets an OAuth client request structured authorization details instead of relying only on broad scopes. For AI agents, the useful pattern carries action, resource, location, and transaction limits into a separate per-request policy decision for the authenticated LLM route.

ai-agentsoauthidentity-and-authorizationpolicy-enforcementllm-security
Read post →

AI Gateway Architecture: The Components That Sit Between an Enterprise Caller and an LLM Endpoint

An AI gateway architecture has six core components: TLS termination, identity binding, request inspection, policy evaluation, the model router, and the audit record emitter. Each component is a placement decision that ties to a regulatory obligation or an operational property. This piece walks through the components, the placement decisions, and how the gateway integrates with the corporate IdP and the SIEM.

ai-gateway-architectureai-gatewayai-securityinline-enforcementaudit-logs
Read post →

Okta AI Gateway Integration: Validate Identity Before LLM Routing

An Okta AI gateway integration should validate the caller access token, derive a small policy context from trusted claims, and make an authorization decision before sending an HTTP request to an LLM provider. This guide defines the request path, claim mapping, denial behavior, evidence record, and the boundary that keeps tool and business API authorization with their resource servers.

oktaai-agentsai-gatewayidentity-and-authorizationpolicy-enforcementarchitecture
Read post →