Cloud Providers

Cloud operations teams ask an assistant to explain an incident, read a runbook, or draft a remediation plan. The danger is not the answer. The danger is an assistant that reaches a live control plane while a shift is under pressure. This starter draws that line before the model is called, and it screens what comes back.

Injection screening runs first, so a crafted instruction cannot disable the checks behind it. Masking then hides values that look like API keys or session identifiers. A block list refuses any prompt that carries the words private key, secret or token.

The tool gate is the part that matters most here. It permits document search and nothing else. Shell access, code execution and database writes are denied by name, so the assistant can read your material but cannot change a running system. Requested tool calls are analyzed on the gateway itself, which keeps command arguments out of a third-party service.

On the return path, an answer must reach a combined score of 0.84, with relevance at 0.78 and accuracy at 0.80. Replies under 120 characters or shorter than two sentences fail the same check and are refused. The review step ships empty, so nothing escalates until you turn it on and name the reviewers. Audit marking records that a request was evaluated, and it never blocks or retains anything by itself.

Identity is not part of this chain. No role check runs, so your own service and your gateway credentials still decide who may ask. Evidence from these checks supports SOC 2 Type II and ISO 27001 questions about tool restriction. It also helps show monitoring readiness under NIS2. Change approval, production safety and personal data duties under GDPR stay with your team.