# A cloud security baseline for small teams: 20 controls

> Twenty cloud security controls for a small team, in the order to do them: identity, keys, secrets, backups, exposure, patching, logging and AI workloads.

- Published: 2026-03-24; last updated 2026-11-14
- Author: the WeStack AI team
- Subject: Cybersecurity and AI governance
- Address: https://westack.ai/blog/cloud-security-baseline-for-small-teams

> In our experience, cloud incidents at small companies begin with a stolen credential, a resource nobody meant to make public, or software nobody patched, and twenty unglamorous controls prevent most of them. Do identity first, then keys and secrets, then backups, exposure, patching, logging and an incident plan, and finish with the two that AI workloads add. A small team can put the whole baseline in place in about a month without a security hire.

## What is a baseline, and which ones are worth borrowing?

A baseline is the minimum every account and workload must meet. It is short enough to check and strict enough that an exception needs a reason.

You do not have to invent one. Three kinds of published guidance cover the ground.

- **The CIS Controls.** Their first tier, [Implementation Group 1](https://www.cisecurity.org/controls/implementation-groups/ig1), is described as essential cyber hygiene: 56 safeguards meant for small and medium organisations with limited security expertise.
- **The CIS Benchmarks.** These are [prescriptive configuration recommendations](https://www.cisecurity.org/cis-benchmarks), with a foundations benchmark each for AWS, Azure and Google Cloud. They tell you which settings to change.
- **The providers' own frameworks.** The [AWS security pillar](https://docs.aws.amazon.com/wellarchitected/latest/security-pillar/welcome.html), Azure's [security checklist](https://learn.microsoft.com/en-us/azure/well-architected/security/checklist), whose first item is to establish a baseline, and Google Cloud's [security, privacy and compliance pillar](https://docs.cloud.google.com/architecture/framework/security), which now includes a principle on using AI securely.

They agree with each other on nearly everything. What they lack is an order. A team of twelve with no security specialist needs to know what to do on Monday and what can wait until week three. This article is our ordering.

Keep one company in mind: twelve people, a SaaS product in a single cloud account, a Postgres database, object storage, a few containers, a CI pipeline, and since the spring an LLM feature that calls a hosted model. Three engineers have administrator access. Nobody owns security.

## Why this order?

We rank controls by how often the failure happens, how bad it is, and how little effort the fix takes.

Credentials come first because that is where cloud incidents usually start. In our experience an attacker signs in far more often than they break in. Backups come early for a different reason: they are the only control that still helps when everything else has failed. Exposure and patching follow, because they shrink what can be reached from outside.

Detection comes fourth, and that is deliberate. An alert is worth something only if a person reads it, and a small team has very little attention to spare. Cut down what can go wrong first, then watch what remains.

## The twenty controls, in order

### Identity (1 to 5)

1. **Single sign-on for every person.** One identity provider for the cloud console, the code host and the CI system. No local users. Offboarding becomes one action.
2. **Phishing-resistant MFA.** Passkeys or hardware security keys, enforced at the identity provider. Codes by text message are better than nothing, and not by much.
3. **Root and owner credentials locked away.** Their own MFA device, no access keys, used only for the few tasks that require them, with an alert on every sign-in.
4. **No long-lived access keys.** People get temporary credentials through SSO. Workloads get roles. The pipeline authenticates through OIDC federation. Delete every key nobody can explain.
5. **Production in its own account, with least privilege.** A separate account, subscription or project. Everyday access is read-only, changes go through the pipeline, and elevated access is time-limited and logged.

### Secrets and data (6 to 10)

6. **A secrets manager, plus scanning for leaks.** Nothing in repositories, container images or chat. Turn on secret scanning with push protection in the code host.
7. **Encryption at rest and in transit.** Mostly a default now. Check older resources, and use TLS on internal database connections too.
8. **Public storage blocked at the account level.** Use the organisation-wide setting that stops any bucket becoming public. Serve public files through a CDN.
9. **Automated backups, copied out of reach.** Database and object storage, copied to a second account with deletion protection, so one stolen credential cannot remove both.
10. **A restore you have actually done.** Time it, write the number down, repeat every quarter.

### Exposure and software (11 to 14)

11. **No public databases or admin ports.** Databases, caches and queues live in private networks. No SSH or RDP open to the internet. Reach machines through the provider's identity-based session access.
12. **A list of what faces the internet.** Build it from the provider's access-analysis tooling and a monthly external scan. Every entry has an owner and a reason.
13. **Patching you do not have to remember.** Prefer managed services, which the provider patches. Turn on automatic operating-system updates and rebuild container images from current base images on a schedule.
14. **Dependency and image scanning in the pipeline.** Automatic update pull requests, and a rule for how quickly critical findings are fixed.

### Detection and response (15 to 18)

15. **An audit log nobody can edit.** Organisation-wide, all regions, delivered to a separate account with write-once retention.
16. **The provider's posture and threat detection, switched on.** AWS Security Hub and GuardDuty, Microsoft Defender for Cloud, or Google's Security Command Center. They can check your accounts against the CIS benchmark. Treat high-severity findings as bugs with an owner.
17. **A short list of alerts that a named person reads.** Root sign-in, a new access key or user, a resource made public, a change to audit logging, a spending anomaly. Send them to whoever is on call, not to a channel everyone has muted.
18. **A one-page incident plan, rehearsed once.** Who decides, how to revoke credentials and isolate a system, whom to call (provider support, lawyer, insurer, customers), and where notes are kept if the main systems are down.

### AI workloads (19 and 20)

19. **Model credentials treated as production credentials.** A key per environment, in the secrets manager, with a spending cap. Where the model is hosted by your cloud provider, use a role, and restrict which models and regions it may invoke.
20. **Data in prompts and GPU capacity under control.** Decide what may go into prompts, protect invocation logs as you protect the database, set GPU quotas to what you use, and never expose an inference server, notebook or vector database without authentication.

## What changes when you add AI workloads?

The first eighteen controls still apply. AI adds a few new ways to lose money and data.

**Keys that turn into a bill.** A database credential is worth whatever the data is worth. A model credential is worth cash from the first minute. In May 2024 Sysdig's threat research team described attackers who took cloud credentials from a vulnerable web application and used them to reach [hosted LLM services, a pattern they named LLMjacking](https://www.sysdig.com/blog/llmjacking-stolen-cloud-credentials-used-in-new-ai-attack). The aim was to sell access while the victim paid, and Sysdig's worst-case estimate of the cost was more than $46,000 a day. Hard caps, narrow roles and a usage alert are the defence.

**Data in prompts and logs.** Prompts carry customer data to the model provider, and invocation logs keep a copy. Both need a decision about what is allowed and how long it stays. We map the options in [data privacy with LLMs](https://westack.ai/blog/llm-data-privacy-and-pii).

**GPU instances as targets.** Expensive compute is worth stealing, whether for mining or for running someone else's workloads. Set quotas to zero in regions you do not use, alert when a GPU instance type is launched, and put a budget on the account. The cost side is in [FinOps for AI](https://westack.ai/blog/finops-for-ai).

**New services with weak defaults.** Self-hosted inference servers, notebooks, vector databases and experiment trackers tend to be stood up from a tutorial, with authentication left for later. Put them in a private network behind single sign-on from the first day.

**Agents with cloud permissions.** An [AI agent](https://westack.ai/services/ai-agents) that can call cloud APIs is a new principal and needs its own narrow role. A developer's key will not do. That design is the subject of [identity and access for AI agents](https://westack.ai/blog/identity-and-access-for-ai-agents). The application-level risks, such as prompt injection and unsafe output, have their own list in our [LLM application security checklist](https://westack.ai/blog/llm-application-security-checklist).

## A checklist by week for the first month

| Week | Controls | Done when |
| --- | --- | --- |
| 1 | 1 to 5: identity | Nobody signs in without SSO and MFA, no access key predates the week, and production is a separate account |
| 2 | 6 to 10: secrets and data | Secret scanning blocks pushes, public storage is blocked, and one restore has been timed |
| 3 | 11 to 14: exposure and software | Only the load balancer and CDN face the internet, and scans run on every build |
| 4 | 15 to 20: detection, response, AI | Five alerts reach a named person, the incident page exists, and model keys have caps |

For a single-account company of this size, a month is realistic for one engineer giving it a good part of each week, with a colleague to review changes.

Week one is the disruptive one. People enrol new sign-in methods, scripts that relied on old keys stop working, and moving production into its own account takes planning. Announce it, pick a quiet week, and do not let it slide, because every later control assumes it.

Twelve people could split the work across engineers. We would still name one owner, because a baseline owned by everyone is checked by no one.

## What breaks, and how do you know it is working?

Four things erode a baseline.

- **Exceptions.** The bucket made public 'for the demo', the key created 'just for this migration'. Give every exception an expiry date and an owner.
- **Alert fatigue.** If item 17 grows from five alerts to fifty, people stop reading. Add an alert only when you can say what the reader should do.
- **Drift.** Settings changed by hand in the console quietly undo the work. Put the baseline in infrastructure code and in organisation policies that prevent a bad configuration, not merely report it. We cover that in [cloud management for growing teams](https://westack.ai/blog/cloud-management-for-growing-teams).
- **An untested break-glass.** The root credentials are locked away so well that nobody can use them in an emergency. Try the procedure once a year.

You can tell it is working from six numbers, read monthly:

1. Long-lived access keys: zero.
2. People without phishing-resistant MFA: zero.
3. Internet-facing resources: a known list, each with an owner.
4. Open high-severity findings, and the age of the oldest.
5. Time to restore the database, measured within the last quarter.
6. Time from an alert firing to a person looking at it.

Be clear about what the baseline does not do. It will not find an authorisation bug in your own code, and it will not stop a determined attacker who targets you specifically. It is a floor. It is also most of the evidence a SOC 2 or ISO 27001 audit asks for, which is a good reason to write things down as you go.

## Where to start on Monday

Do three things before the end of the first day. Enforce single sign-on with MFA for the cloud console. List every access key in the account and delete the ones without an owner. Start a restore of last night's database backup into a scratch environment and time it.

If you have an AI feature, add a fourth: put a spending cap on every model key. It is a short job, and it covers the failure that costs the most per hour.

Then put weeks two to four in the calendar with a name against each. If you would like the baseline set up or reviewed, this is what our [cybersecurity](https://westack.ai/services/cybersecurity-ai-governance) and [cloud and DevOps](https://westack.ai/services/cloud) teams do.

## Frequently asked questions

### What is a cloud security baseline?

A baseline is the minimum set of controls that every account and workload must meet, written down and checked regularly. Recognised starting points are the CIS Controls, the CIS Benchmarks for each cloud provider, and the security pillars of the AWS, Azure and Google Cloud well-architected frameworks. A small team needs a short version it will actually finish.

### What should a small team secure first in the cloud?

Identity. Put everyone behind single sign-on with phishing-resistant multi-factor authentication, lock away the root or owner credentials, and remove long-lived access keys. Then make sure backups exist in a separate account and that one has been restored. Those steps address account takeover and unrecoverable data loss, the two failures that are hardest to come back from.

### Do small companies need the CIS Benchmarks?

They are worth using as a checking tool, not as a project. The foundations benchmark for your provider lists specific settings, and the provider's own posture tool can score your accounts against it automatically. Fix the high-severity findings and the ones that match this baseline, and leave the long tail until a customer or auditor asks.

### How does AI change cloud security for a small team?

It adds credentials that convert directly into cost, such as model API keys and cloud roles allowed to invoke hosted models. It puts personal and confidential data into prompts and logs. And it brings GPU instances and new services such as inference servers, notebooks and vector databases, which are attractive targets and are often deployed without authentication.
