Bitdefender GravityZone AIVC Detects Local Model Runtimes and MCP Servers Missed by Standard Inventories

The module surfaces four exposure classes: browser-based hosted assistants, editor-embedded coding tools, local runtimes executing directly on endpoints, and configured connections to databases or APIs. Discovery occurs via the BEST agent metadata collection, which assigns each finding to one of 17 functional categories and flags unapproved shadow AI regardless of domain lists. This directly addresses gaps where traditional software inventories fail on extensions and configuration entries.

Local model runtimes and MCP servers represent the highest blind spot because they bypass network controls entirely and often arrive without registry entries. The 70,000 skills extend agent permissions to external systems, expanding attack surface beyond the agents themselves. This pattern aligns with documented increases in unsanctioned AI use following releases such as Anthropic Claude 3.5 and open-source agent frameworks in 2024-2025.

GravityZone console presents two inventory views plus policy connectors that map findings to approval workflows. Security teams can now filter by data-handling risk categories while development leads isolate coding-tool exposure. Independent verification of similar capabilities appears in endpoint telemetry reports from CrowdStrike and Microsoft Defender for Endpoint, confirming the same inventory blind spots.

Next steps include integration with existing DLP policies and expansion beyond Windows endpoints. Organizations should validate whether AIVC findings correlate with actual data exfiltration attempts within the next quarter.