Just as software teams maintain a Software Bill of Materials (SBOM) for their dependencies, AI-first security teams need an AI BOM: a complete, auditable registry of every model in use, its version, its access controls, and its risk profile.
A centralised register of every model: OpenAI GPT-4o, Anthropic Claude, Google Gemini, Azure OpenAI endpoints, self-hosted Llama variants, fine-tuned models — all in one place.
Models change. Track which model version each application is using, when it was last updated, and whether a security advisory has been issued for any deployed version.
For each model, see which teams own it, which applications call it, and which users have access. Spot over-permissioned access with one view.
Each model receives a risk score based on: data sensitivity of its inputs, access permissions, compliance certifications, vendor security posture, and known vulnerabilities.
Kovral builds the inventory automatically from proxy traffic, then enriches it with metadata from your cloud provider accounts and manual registration for models Kovral can't observe directly.
Every model API call through the Kovral proxy is automatically catalogued. New models appear in the inventory within minutes of first use.
Connect your AWS, Azure, or GCP account (read-only) and Kovral scans for AI service resources: SageMaker endpoints, Azure OpenAI deployments, Vertex AI models, and Bedrock model access grants.
Models not reachable via proxy or cloud scanning (air-gapped, embedded, or third-party) can be manually registered with metadata. Manual entries integrate with all governance workflows.
Model access patterns are monitored continuously. Unusual usage — a model called by an unexpected application, a spike in usage volume, a new geographic access point — triggers an alert.
