← All posts

Platform & Architecture

338 posts on platform & architecture.

Bedrock AgentCore Audit Logs: Build a Four-Record Evidence Map

Amazon Bedrock AgentCore emits OpenTelemetry-compatible traces, CloudWatch metrics and logs, and CloudTrail events for supported capabilities. Those records answer different questions about agent execution, service administration, and Gateway invocation. This guide builds a four-record evidence map across telemetry, CloudTrail, IAM, and model-route authorization, with explicit collection tests and ownership boundaries for AgentCore workloads.

agentic-aiai-securityauditforensic-auditidentity-and-authorizationcloud-security
Read post →

Fireworks AI Security Needs a Route Acceptance Matrix

Fireworks offers shared serverless inference and dedicated GPU deployments with service accounts, model policy, regional controls, audit logs, and deployment metrics. A production review still needs a route acceptance matrix that binds each authenticated HTTP request to a principal, endpoint, model, region, data class, and evidence record.

ai-securityllm-securitycloud-securityidentity-and-authorizationpolicy-enforcement
Read post →

Cerebras Inference Audit Logs Separate Requests From Console Actions

Cerebras Inference Cloud exposes request logs for API calls and a separate admin-only audit log for console actions. Projects and API credentials add useful attribution for workloads, quotas, and account changes. This guide maps each record to the question it can answer and defines the extra evidence needed for an authenticated model request.

ai-securityllm-securityauditforensic-auditidentity-and-authorizationarchitecture
Read post →

Instruction Hierarchy for LLMs Needs a Privileged Regression Pack

An instruction hierarchy gives higher-authority instructions precedence when system, developer, user, and untrusted content conflict. This guide turns that model behavior into a privileged-instruction regression pack that records roles and trust labels, expected precedence and response, principal and policy, plus model version.

llm-securityprompt-injectionai-securitypolicy-enforcementidentity-and-authorization
Read post →

Haystack Security Starts With the Trusted Execution Boundary

Haystack expects applications to validate input, trust executable pipeline definitions, constrain network access, and defend prompts that contain retrieved content. This guide turns those assumptions into a trusted-execution and generator-route checklist with named owners, bypass tests, and request evidence.

ai-securityllm-securityarchitectureprompt-injectionpolicy-enforcement
Read post →

Flowise Security: A Deployment Acceptance Test for Production

Flowise security starts with separate decisions for application access and chatflow access. A production acceptance test should also prove the public-flow posture, IP-based rate limiting through the real proxy chain, caller identity propagation, and the outbound model route. This guide assigns a pass condition and retained evidence to each boundary, while keeping host, credential, flow, and separately routed LLM controls distinct.

ai-securityllm-securityidentity-and-authorizationpolicy-enforcementarchitecture
Read post →

MCP Streamable HTTP Security: A Reverse-Proxy Acceptance Test

The 2026-07-28 MCP Streamable HTTP transport sends each JSON-RPC request as its own POST and permits either a JSON response or a request-scoped SSE stream. It also exposes protocol and method metadata in required headers, plus target metadata when applicable. The server must verify those headers against the body. A production acceptance test should cover header-body mismatches, unsupported versions, OAuth audience checks, stream buffering, cancellation, and downgrade behavior.

ai-securityagentic-aillm-securityidentity-and-authorizationpolicy-enforcementarchitecture
Read post →

Knowledge Graph LLM Security Starts at Context Assembly

A knowledge graph can add entities, relationships, claims, community reports and source text to one LLM context window. Security teams need to preserve authorization through that assembly process, then inspect the completed prompt before it reaches the model. This guide defines a context manifest, route policy and decision record for customer-controlled GraphRAG applications while keeping graph construction, source permissions and model traffic with their proper control owners.

ai-securityllm-securityidentity-and-authorizationpolicy-enforcementforensic-audit
Read post →

Format-Preserving Encryption for LLM Traffic Has a Narrow Job

Format-preserving encryption replaces a structured value with ciphertext that keeps the original length and alphabet. That makes it useful for selected legacy fields crossing an LLM request, such as fixed-width identifiers. It is a poor default for prose, needs careful domain design plus controlled handling of encryption secrets and tweaks, and leaves authorization, prompt context, output validation, and re-identification as separate controls.

ai-securitydata-loss-preventionllm-securitypolicy-enforcementarchitecture
Read post →

Implement OpenTelemetry GenAI Conventions at an AI Gateway

OpenTelemetry GenAI semantic conventions give AI gateways shared names for operations, providers, models, errors, spans, and duration metrics. A production implementation still needs a gateway-owned policy span, careful treatment of prompt content, stable correlation with audit evidence, and a migration plan because many GenAI fields remain in development.

ai-securityarchitecturecloud-securitypolicy-enforcementforensic-audit
Read post →

Apple Private Cloud Compute Changes the Cloud AI Trust Contract

Apple Private Cloud Compute processes selected Apple Intelligence requests on attested cloud nodes. Devices encrypt each request to validated nodes, releases appear in a public transparency log, and the service limits privileged access and retention. Enterprise teams still need a separate decision about which users may send which data through AI routes they control.

ai-securitycloud-securityarchitecturezero-trustidentity-and-authorization
Read post →

Qwen3Guard Deployment: Classifier Signals, Streaming, and Policy Boundaries

Qwen3Guard provides generative and streaming safety classifiers for prompts and model responses. Production teams still need to define the action taken for each label, bind that action to an authenticated caller and route, test failure behavior, and retain evidence that separates classification from authorization.

ai-securityllm-securitypolicy-enforcementinline-enforcementarchitecture
Read post →