Share with your CISO
Snyk’s State of Agentic AI Adoption report, drawn from 3,044 enterprise environments and 1.39 million code repositories, finds that the average enterprise AI footprint is roughly three times larger than its model inventory suggests. Nearly half of organizations using AI carry zero declared models in their repositories, running entirely on third-party services and packages. Meanwhile 46.9% have adopted agentic architectures combining AI agents, MCP servers, or both, and 77.4% of AI packages and tools originate from external sources.
What this means for your business
The story most likely to catch a security leader off guard here is not the model count but the supply chain underneath it. If your AI governance program is built around a model registry, you are auditing roughly one-third of your actual exposure. The 77.4% external-sourcing figure means most of the code executing decisions in your AI systems was written by someone outside your organization, and the roughly 50% of organizations that cannot link deployed models to their training datasets are flying blind on both incident investigation and compliance attestation.
Snyk sells developer security tooling and has an obvious interest in making the AI attack surface look wide, but the underlying data point holds regardless of who collected it. Enterprises have always underestimated their software supply chain risk, and AI stacks follow exactly the same pattern, only faster. The addition of MCP servers, which act as integration layers letting AI agents reach into enterprise data, applications, and external tools, multiplies that surface in a way that traditional software composition analysis was never designed to catch. A framework dependency that quietly gains write access to a production database is a qualitatively different risk than a vulnerable npm package.
The falsification condition for concern here is straightforward. If your organization has mapped not just models but every orchestration framework, vector database, retrieval pipeline, and MCP server in production, and you can trace the training lineage on the models that touch regulated data, this report describes someone else’s problem. Most enterprises cannot do that today. The governance gap is the budget conversation, not the model count.
Concept deep-dive: Model Context Protocol (MCP)
MCP is a standardized interface, originally introduced by Anthropic, that lets AI agents connect to external data sources, applications, and tools in a structured way, think of it as a USB standard for AI integrations. Instead of each agent requiring custom connectors, MCP servers expose capabilities that any compatible agent can call. The business implication is that a single MCP server can grant an AI agent broad access to enterprise systems, making it both a productivity multiplier and a high-value target for compromise or misconfiguration.
Based on reporting from Your enterprise AI footprint is about three times bigger than your model list, originally published 2026-08-05 00:00:00.
