We recently reviewed our cloud accounts for unauthorized AI deployments and found several Azure OpenAI instances that had never been approved, a self-hosted model with overly broad access to internal data, and a Replit application sending customer information to an external large language model without the security team knowing.
Traditional cloud security tools have not been great at finding this kind of activity. Most CNAPP products focus on infrastructure posture and vulnerabilities, but they do not always identify AI models, trace their data access, or detect sensitive information being sent to external model providers.
We have been evaluating AI Security Posture Management tools, but the market seems split. Some platforms offer AI-SPM features only as part of a larger suite and expect you to know what to search for, while some endpoint-focused vendors have cloud modules that missed several of our shadow AI deployments.
What processes or tools are you using to discover unauthorized models, AI applications, and data pipelines before they turn into a security incident?
5 Answers
For self-hosted models, focus on identity and data access. Review the managed identities, service accounts, RBAC assignments, storage permissions, and data-plane roles attached to the model's compute. A model with broad access to sensitive datasets may be a bigger risk than an unapproved endpoint by itself. Pairing CNAPP with DSPM can help map sensitive data to the identities, applications, and AI systems that can reach it.
I would separate discovery from prevention. For Azure, use Azure Resource Graph across every subscription to search for Cognitive Services and OpenAI resource types, then review Defender for Cloud's AI posture findings. Azure Policy can prevent new deployments outside approved subscriptions, but it will not tell you what is already running.
A practical control is to deploy approved models through Terraform and require an approval workflow for anything outside the standard catalog. For monitoring, compare cloud billing and provider usage with application telemetry; unexplained Azure consumption or a mismatch between billed usage and what monitoring sees can expose hidden deployments. It is not perfect, but it provides a useful detection signal.
Start by treating AI assets as part of the normal cloud inventory rather than maintaining a completely separate list. Identify model endpoints, related compute, service identities, storage connections, and external API destinations, then review those relationships regularly. Discovery has to cover both resources inside the tenant and data leaving through endpoints, SaaS tools, and developer platforms.
The Replit example is a different problem because the application is outside your cloud tenant. That is an outbound data-flow issue, so you need network and firewall logs, NSG flow logs, proxy telemetry, or a secure web gateway that can identify traffic to external LLM endpoints. Cloud inventory tools generally cannot see those applications.
Related Questions
Biggest Problem With Suno AI Audio
How to Build a Custom GPT Journalist That Posts Directly to WordPress