Quick answer

Shadow AI is any AI tool, model, service or API used or deployed without the knowledge and approval of security and IT. It includes employees pasting company data into consumer chatbots, teams creating cloud AI resources outside governance, AI features quietly switched on in SaaS tools, and personal API keys in production code. You find it by inventorying cloud AI resources, reviewing network, identity and expense data, and giving people a sanctioned alternative.

Key takeaways

  • Shadow AI isn't only employees using chatbots. Forgotten cloud AI resources and stray API keys are often the bigger risk.
  • Every major cloud can list its AI resources in one query. Run it across all accounts, not just the ones you know.
  • Network, identity and expense data reveal AI usage that cloud inventories miss.
  • Blocking alone drives AI further underground. Pair discovery with an approved, secure option.

Ask a security team how many AI services their company runs and you'll get a number. Run an inventory and you'll get a bigger one. The difference is shadow AI: AI tools, models and services in use without security's knowledge or approval.

It's the first thing we look for in every audit, because it's control domain D01 in our certification, and every other control depends on it. You can't apply guardrails, private networking or logging to a model nobody knows exists.

Where shadow AI hides

TypeExampleMain risk
Consumer AI useEmployees pasting code, contracts or customer data into free chatbotsData leaves the company under consumer terms
Unmanaged cloud AIAzure OpenAI, Bedrock or Vertex AI resources in sandbox or project accountsPublic endpoints, API keys, no logging, real data
Stray API keysPersonal OpenAI or Anthropic keys in apps, scripts or notebooks, paid on a credit cardUnmonitored data flows; leaked keys
SaaS AI featuresAI assistants switched on in your CRM, help desk or document toolsVendors processing your data in new ways
Self-hosted modelsA model server on a GPU VM "just for testing"Unauthenticated inference endpoints exposed to the internet
Browser extensions & OAuth appsAI writing assistants granted access to email and documentsBroad, persistent access to mailboxes and files

Find it in your cloud accounts

Run these across every account, subscription and project — especially the ones labeled sandbox, dev or personal.

Microsoft Azure

In Azure Resource Graph Explorer, scoped to all subscriptions:

resources
| where type in~ ('microsoft.cognitiveservices/accounts',
                  'microsoft.machinelearningservices/workspaces',
                  'microsoft.search/searchservices')
| project name, type, kind, subscriptionId, resourceGroup, location

Amazon Web Services

Look for Bedrock activity in CloudTrail in each account and region:

aws cloudtrail lookup-events \
  --lookup-attributes AttributeKey=EventSource,AttributeValue=bedrock.amazonaws.com \
  --max-results 50

Also list SageMaker endpoints (aws sagemaker list-endpoints), and check for GPU instances that might host self-managed models. At scale, use an organization trail with CloudTrail Lake or Athena instead of running this account by account.

Google Cloud

Search Cloud Asset Inventory across the organization:

gcloud asset search-all-resources \
  --scope=organizations/ORGANIZATION_ID \
  --asset-types='aiplatform.googleapis.com.*'

Also check which projects have generativelanguage.googleapis.com (the Gemini Developer API) enabled — a common sign of API-key-based usage outside Vertex AI.

Find it outside the cloud console

  • Network and DNS logs: traffic to api.openai.com, api.anthropic.com, generativelanguage.googleapis.com and other model APIs from servers that shouldn't be calling them.
  • Secure web gateway / CASB: which AI web apps employees use, how often, and whether files are being uploaded.
  • Identity provider: OAuth app consents in Microsoft Entra ID or Google Workspace that grant AI tools access to mail, files or calendars.
  • Code repositories: secret scanning for AI provider API keys, and dependency scans for AI SDKs in services nobody registered.
  • Expense reports and card statements: subscriptions to AI tools and API usage bills. Finance often knows about shadow AI before security does.
  • SaaS admin consoles: AI features enabled in the tools you already pay for.

What to do with what you find

  1. Triage by data and exposure. A public endpoint with customer data is an incident. A sandbox model with test data is a ticket.
  2. Bring keepers under governance. Assign an owner, move it into a governed account, and apply your platform baseline — Azure, AWS or Google Cloud.
  3. Retire the rest — and revoke the keys, not just the resources.
  4. Offer a paved road. Give people an approved AI assistant and an approved way to build with models. Shadow AI shrinks fastest when the sanctioned option is easier than the workaround.
  5. Make it continuous. Guardrails like Azure Policy, AWS service control policies and Google Cloud Organization Policy can restrict where AI resources can be created. Re-run discovery every quarter.

Start small: run the three cloud queries above this week. Most organizations find at least one AI resource nobody on the security team knew about.

A complete, current AI inventory is the first control we verify in every CertiSafe AI Secure AI Deployment audit, and we re-run discovery every quarter to catch what's new.

Frequently asked questions

What is an example of shadow AI?

A product team creates an Azure OpenAI resource in a sandbox subscription to prototype a feature, connects it to a copy of customer data, uses API keys, leaves public network access on, and ships it to production without a security review. Or an employee pastes a confidential contract into a free consumer chatbot to summarize it. Both are shadow AI.

Why is shadow AI a security risk?

Because nothing about it has been reviewed: data may be sent to providers with unknown retention and training terms, cloud resources may be publicly reachable with long-lived keys, nobody is monitoring usage, and the AI may have access to data or actions it shouldn't. You can't secure, log or respond to incidents in systems you don't know exist.

Should companies block generative AI tools to stop shadow AI?

Blocking everything usually backfires — people switch to personal devices and accounts where you have no visibility at all. A better approach is to provide approved, secure AI tools that meet people's needs, block or restrict only the highest-risk services, and make it easy to request new tools through a fast review process.

Get certified against the leaks covered in this article

CertiSafe AI tests your live AI features every quarter and gives you a verifiable stamp for your website.

Schedule a Meeting