# AI Security for Cloud Teams Checklist

Use this checklist before scaling AI workloads, model gateways, agent sandboxes, notebooks, SaaS connectors, cloud service accounts, or AI-assisted cloud remediation.

## 1. AI Asset Inventory

- [ ] Inventory AI workloads, model endpoints, vector stores, notebooks, data buckets, agent runtimes, and SaaS connectors.
- [ ] Map owners, environments, data classes, cloud accounts, regions, and production dependencies.
- [ ] Include AI systems in CNAPP, CSPM, asset inventory, and exposure management workflows.
- [ ] Detect shadow AI infrastructure created outside approved platform paths.

## 2. Identity and Access

- [ ] Bind cloud credentials to expected repositories, branches, workflows, workloads, and environments.
- [ ] Prefer short-lived credentials and workload identity over static keys.
- [ ] Separate human, service account, workload, and AI agent identities.
- [ ] Apply least privilege to model gateways, data stores, build systems, and remediation tools.

## 3. Model Gateway and Key Control

- [ ] Route model usage through approved gateways where possible.
- [ ] Track team ownership, model provider, model alias, token budget, cost, and policy decisions.
- [ ] Restrict provider keys, virtual keys, and model access by workflow and data sensitivity.
- [ ] Rotate leaked model keys and review logs for abnormal spend or provider routing.

## 4. Network and Egress Control

- [ ] Restrict egress from agent sandboxes, build runners, notebooks, and experimentation environments.
- [ ] Block unnecessary internet access from sensitive AI workloads.
- [ ] Monitor outbound traffic to package registries, paste sites, unknown APIs, and unexpected model providers.
- [ ] Review MCP servers and connectors that bridge local tools, cloud APIs, and SaaS systems.

## 5. Containment and Recovery

- [ ] Prepare runbooks for compromised service accounts, leaked model keys, malicious automation, and cloud control-plane abuse.
- [ ] Test revocation of model keys, cloud roles, OIDC trust policies, deploy keys, and SaaS app grants.
- [ ] Keep clean-room recovery paths for identity, CI/CD, cloud admin, and production restore operations.
- [ ] Verify backups, logs, and alternate communications before a breach.

## 6. Metrics

- [ ] Track AI workload inventory coverage, overprivileged identities, unmanaged model keys, blocked egress events, and incident containment time.
- [ ] Review cloud AI security posture at least monthly while adoption is growing.

Related HackWednesday pages:

- https://hackwednesday.com/ai-security-for-cloud-teams
- https://hackwednesday.com/corporate-network-defense-blueprint
- https://hackwednesday.com/resources/wiz-ai-security-skills-for-cloud-teams
- https://hackwednesday.com/zero-trust-ai-agent-security