AI Infrastructure and Deployment Security

GAIPS · 25 questions

• A city deploys a permitting LLM on a managed cloud AI service. Security reviews how municipal prompts leave the city’s network toward the provider. What control focus is most appropriate?
• County public health runs open-weight models on GPU servers that share an unsegmented VLAN with unrelated clinic systems. What infrastructure change best reduces lateral risk?
• Transit exposes rider-assistant model endpoints behind a public load balancer without WAF rules tuned for prompt abuse patterns. What edge control is missing?
• A housing authority stores fine-tuned model weights in an object bucket with worldwide read ACLs “for convenience.” What risk is the team introducing?
• Library GenAI pods run as root and mount hostPath volumes to node scratch disks that may hold resident prompt fragments. What hardening priority applies?
• Emergency management compares a managed cloud AI service with self-hosted models for continuity if a regional cloud outage hits. What decision factor belongs in that hosting review?
• A school-district Kubernetes cluster colocates student GenAI tutoring apps with unrelated finance workloads on shared nodes. What isolation concern should platform security raise?
• Water-utility inference nodes lack disk encryption, and crash dumps may contain prompt fragments from resident service chats. What host control addresses residual GenAI data?
• City SOC finds unmanaged Ollama-style local model runtimes on analyst laptops processing case notes. What governance action fits infrastructure security?
• Permitting GPU nodes share a single cloud IAM role that can also modify IAM policies. What identity principle is violated for GenAI compute?
• County open-data GenAI caches personalized answers on a CDN keyed only by client IP. What cache-design risk should security call out?
• Appsec reviews whether model-serving admin consoles for the city’s inference platform enforce SSO and MFA. What control is under review?
• Transit’s DR plan replicates vector databases and model artifacts to a second region without reviewing encryption-in-transit on those paths. What gap exists?
• Public-health GenAI autoscaling launches new nodes that briefly expose ephemeral debug ports. What bootstrap hygiene is needed?
• Court AI summarizer VMs allow unrestricted outbound internet “to fetch citations.” What egress control should security require?
• A housing GenAI Kubernetes namespace has no network policies, so any pod can reach the vector database. What control is missing?
• The city evaluates confidential-computing options for GenAI processing of sealed juvenile records. What capability is being considered?
• Parks stores model-hub tokens in CI variables readable by all project members deploying GenAI. What credential-handling fix is needed?
• County IT considers GPU cloud spot instances for batch GenAI summarization against a strict data-residency policy. What hosting alignment check matters most?
• Overnight, the 311 production model endpoint still points at a shared “dev” inference cluster. What deployment hygiene failed?
• Security scans find a known CVE in the city’s self-hosted LLM inference gateway software. What infrastructure practice is indicated?
• Emergency operations wants air-gapped model updates delivered with signed USB transfer procedures. What update-channel property is essential?
• A clerk SaaS GenAI add-on processes court filings in a foreign region contrary to local records policy. What hosting validation failed?
• Transit exposes Jupyter on the same GPU nodes that serve production rider-assistant models. What separation is required?
• The city CISO requires logs of model ID, endpoint, and caller identity for every civic inference call. What infrastructure capability is being mandated?