nsg-audit

Imported from microsoft/sre-agent (labs/zava-learning/sre-config/agent-config/skills/nsg-audit/SKILL.md). Install upstream with npx skills add microsoft/sre-agent --skill nsg-audit. Copyright stays with the author.

Zava Learning — Network Security Group (NSG) & Connectivity-Path Audit

Resource Group: @@RG@@. Read-only audit — never create, modify, or delete an NSG, rule, route, or firewall. This is a weekly posture review of the lab's network exposure and the integrity of its connectivity paths, not an incident response.

1. Enumerate the network surface

• List every NSG: az network nsg list -g @@RG@@.
• For each NSG, list its security rules (custom + default): az network nsg rule list -g @@RG@@ --nsg-name <nsg> --include-default.
• Map each NSG to what it protects (az network nsg show -g @@RG@@ -n <nsg> --query "{subnets:subnets[].id, nics:networkInterfaces[].id}"). An NSG attached to no subnet and no NIC is orphaned — itself a finding.

2. Evaluate every inbound (and risky outbound) rule

Flag, with an impact-based severity (see the zava-audit-report severity model):

• SEV1 — any inbound Allow from * / Internet / 0.0.0.0/0 to a management port (22 SSH, 3389 RDP); a blanket Allow *:* * (any source, any port, any protocol); or a rule whose description marks it temporary / break-glass ("temp", "troubleshooting", "INC-####") that is still present.
• SEV2 — inbound Allow from a broad source (large CIDR / Internet) to a data or admin port (e.g. 5432 Postgres, 6379 Redis, 1433, 27017, 8080-808x app-management ports) where the source should be a known subnet / Application Gateway only.
• SEV3 — shadowed rules (a higher-priority rule fully masks a lower one so it can never match), unused / legacy rules (deny or allow referencing a decommissioned CIDR/subnet), overlapping / duplicate rules, priority inversions, and orphaned NSGs (attached to nothing). Distinguish the lab's intended exposure (the Application Gateway public lane ports 8081-8087 reaching the lane apps; the AppGw subnet allowing Internet inbound on 80/443 + those lane ports) from genuinely risky exposure — flag only what a real reviewer would. The lane data ports (Postgres, the reporting-VM subnet) should be reachable only from inside the VNet; the reporting VM has no public path and needs no inbound Internet at all.

3. Trace the full connectivity path (not just NSGs)

A real network audit follows the whole path. Enumerate and assess each layer, and flag misconfigurations:

• UDRs / route tables — az network route-table list -g @@RG@@ and az network route-table route list. Flag routes that bypass a firewall / NVA (e.g. 0.0.0.0/0 -> Internet where an appliance is expected), black-hole routes (nextHopType=None), routes to decommissioned next hops, and route tables not associated with any subnet.
• Private Endpoints & Private DNS — az network private-endpoint list -g @@RG@@; for data services (Postgres, Key Vault, Storage) check whether access is via Private Endpoint or left on a public endpoint, and whether the matching Private DNS zone + VNet link exist (az network private-dns zone list, ... link vnet list). Flag a data service exposed publicly when a Private Endpoint is expected, or a Private Endpoint with a missing/dangling DNS zone link (which silently breaks name resolution).
• DNS — confirm Private DNS zones resolve to the Private Endpoint NICs and have a VNet link; flag orphaned zones or links to deleted VNets.
• Firewall / NVA — if an Azure Firewall or other NVA is present (az network firewall list -g @@RG@@), check that subnet UDRs actually force egress through it and review overly-permissive network/application rules; if none is present, note that egress is direct and whether that is acceptable for the lab. Where a control reference helps, cite Azure network-security guidance via microsoft-learn_microsoft_docs_search (deny-by-default, just-in-time access, hub-spoke segmentation, Private Link, forced tunneling).

4. Produce the findings table

For every finding output a row: Layer (NSG/UDR/PE/DNS/FW) · Resource · Rule/Route · Source → Dest:Port · Finding · Severity · Recommended fix. Sort SEV1 → SEV3. Also produce a short posture summary (Healthy / Needs attention / At risk, counts by severity, the single most important action) and a prioritized recommendations list.

5. Hand off to the branded report

This skill does not remediate. Pass the posture summary, the findings rows (each with its severity), and the recommendations to the zava-audit-report skill, which renders the single branded, downloadable PowerPoint deck and returns its download link. Surface that link to the operator.