/iac-scan

Scan changed infrastructure/deploy config for security misconfigurations and exposed secrets. Static and read-only — analyzes the files directly.

Auto-run safety envelope: the auto-run path uses only local static linters that need no network and no auth — hadolint, tfsec, checkov (offline/local policy only), kube-score, actionlint, zizmor. That's why it's safe to auto-run, unlike /pentest (an interactive check gated on an authorization confirmation, whose opt-in clearwing modes are networked scanners that bill their own LLM provider — its default in-harness source hunt touches no network but still waits for authorization). Anything that reaches the network — e.g. trivy pulling vulnerability DBs, or checkov fetching remote policies — is opt-in / recommend-only, never part of the auto-run. If only networked tools are available, run the signature checks instead and recommend the networked scan separately.

Hard rules

• Evidence-grounded: file:line, the offending directive, the rule/CWE, and the concrete risk ("container runs as root", "0.0.0.0/0 on port 22", "secret baked into image layer").
• Severity by blast radius: internet-exposed, privilege-escalating, or secret-leaking misconfigs are CRITICAL/HIGH; style/pinning nits are LOW.
• Smallest correct fix, shown as a diff; do not apply it.
• Scope to the changed infra files. Don't re-scan the whole tree.
• Prefer real scanners when present. Local static linters (hadolint, tfsec, kube-score, actionlint, zizmor, offline checkov) are part of the auto-run path; networked scanners (trivy DB pulls, remote-policy checkov) are recommend-only. Fall back to the signature checks below when none are present — and always say which mode ran (which tools, or signature mode).

Trigger (what counts as infra/deploy config)

• Containers: Dockerfile, *.dockerfile, docker-compose*.y*ml, .dockerignore
• Terraform: *.tf, *.tfvars
• Kubernetes / Helm: *.y*ml with kind: under k8s/, manifests/, charts/, helm/, or deploy/
• CI/CD: .github/workflows/**, .gitlab-ci.yml, Jenkinsfile, *.circleci/*
• Web/proxy/cloud: nginx.conf, Caddyfile, *.tf cloud resources, serverless.yml

If none of these changed, say so and stop.

What to check

Docker — non-root USER; pinned base image (no :latest); no secrets in ENV/ARG/layers; no ADD http://…; minimal surface; HEALTHCHECK; no unnecessary EXPOSE; --no-install-recommends/cleaned caches. (hadolint)

docker-compose — no privileged: true; no network_mode: host; never bind /var/run/docker.sock; secrets via secrets: not environment:; pinned image tags.

Terraform — no security-group/firewall 0.0.0.0/0 on sensitive ports (22/3389/DB); no public buckets/blobs; no IAM Action: "*" / Resource: "*"; encryption at rest + in transit enabled; no hardcoded secrets/keys; remote state with locking; logging/versioning on. (tfsec / checkov)

Kubernetes / Helm — runAsNonRoot, drop privileged/allowPrivilegeEscalation; no hostPath/hostNetwork/hostPID; resource limits/requests set; readOnlyRootFilesystem; no default ServiceAccount with broad RBAC / cluster-admin; pinned image digests; NetworkPolicy present; no secrets in plain env. (kube-score / kubesec)

CI/CD (GitHub Actions et al.) — no pull_request_target running untrusted PR code; least-privilege top-level permissions:; third-party actions pinned to a full commit SHA (not a moving tag); no script injection via unsanitized ${{ github.event.* }} in run:; secrets never echoed; no persist-credentials left on when not needed. (actionlint + zizmor)

Web/proxy — TLS enforced; security headers (HSTS, CSP, X-Content-Type-Options); no directory listing; no server tokens leaking versions.

Steps

1. Scope — resolve changed infra files (or $ARGUMENTS). None → stop.
2. Tool pass — for each file type, if the matching scanner is installed, run it scoped to the changed files and capture findings verbatim. Note which tools ran.
3. Signature pass — for file types with no installed scanner, Grep/Read for the patterns above. Mark these Observed (clear match) vs Suspected.
4. Secrets sweep — across all changed infra files, flag high-entropy strings, AKIA…, private keys, tokens, password=/secret= literals. (Complements /defense's app-code secret scan.)
5. Report — group by severity; smallest fix per finding as a diff.

Report format

# IaC scan — <scope> @ <date>

Scanners run: hadolint ✓ / tfsec ✗ (signature mode) / actionlint ✓ …

## CRITICAL / HIGH
- <file:line> — <directive> — <rule/CWE> — <risk> — <fix>

## MEDIUM / LOW
- …

## Secrets
- <file:line> — <kind> — rotate + move to a secret manager

Anti-patterns (do not do)

• Claiming a misconfig without quoting the directive and naming the risk.
• Scanning the whole repo instead of the changed infra files.
• Applying fixes or rewriting manifests — this scans and drafts; the user applies.
• Treating a pinning/style nit as CRITICAL, or burying an open-to-the-world port.

Related commands

• /defense — app-code OWASP/secrets sweep; /iac-scan is its infra counterpart.
• /pentest — deeper security hunt (in-harness pipeline by default; network probing drives clearwing's MCP machinery — authorization-gated — and its hunt/SARIF pipeline runs agent-driven on the session via the bundled bridge, with provider-driven runs billed and ask-first); recommend after /iac-scan finds HIGH/CRITICAL or when a live environment needs probing.
• /cso — threat-model the deployment topology at design time (/write-plan).