# Snyk Finds Enterprise AI Footprints Extend Far Beyond Model Inventories > Source: > Published: 2026-08-05 05:01:56+00:00 # Snyk Finds Enterprise AI Footprints Extend Far Beyond Model Inventories Snyk's 2026 State of Agentic AI Adoption: Volume II says its analysis of more than 3,000 enterprise accounts and 1.39 million repositories found that models represent only part of the deployed AI footprint. Help Net Security reported that 46.9% of organizations using AI had agents, Model Context Protocol servers or both, while counting tools, datasets and retrieval systems made the average footprint about three times larger than model inventories alone. Snyk's second 2026 State of Agentic AI Adoption report argues that a model list is not an enterprise AI inventory. Its official Volume II page says the analysis covers more than 3,000 enterprise accounts and 1.39 million code repositories, with models representing only one part of the software and data assembled around AI applications. Help Net Security, reporting on the study on August 5, gave the more precise sample as **3,044 enterprise environments** and **1.39 million repositories**. It said 46.9% of organizations using AI had adopted architectures built around AI agents, Model Context Protocol (MCP) servers or both. Snyk's own landing page describes agentic architectures as running in 33% of the full enterprise sample. Those percentages use different denominators and should not be treated as contradictory. ### The inventory gap sits around the model The study's central finding is that counting frameworks, agents, MCP servers, retrieval systems, vector databases, datasets and supporting tools makes the average AI footprint roughly three times larger than a model-only inventory. Help Net Security also reported that nearly half of the examined organizations had no declared AI models in their code repositories, even though AI services can enter applications through hosted APIs and third-party packages. The source mix adds a software-supply-chain dimension. According to Help Net Security, third-party software represented **77.4%** of AI packages and tools in the study, while internally developed components represented 22.6%. Proprietary models accounted for 63.8% of identified deployments and open-source models for 32.5%; the remaining share was not specified in the report coverage. ### What security teams can take from the data A practical inventory therefore needs relationships, not just names. Teams need to map which agents call which models, what MCP servers and connectors they can reach, which datasets or vector stores supply context, and where third-party packages enter the path. Permissions, data flows and ownership matter as much as model identity. The figures are telemetry from Snyk-observed environments, not a representative survey of every enterprise, so they should be read as evidence of deployment patterns rather than a universal adoption rate. They do, however, show why model-only governance can miss material parts of an agentic system's attack surface. ## Key Points - 1Snyk's Volume II analysis covers more than 3,000 enterprise accounts and 1.39 million code repositories. - 2Help Net Security reported that 46.9% of AI-using organizations had agents, MCP servers or both, and that full AI footprints averaged about three times model-only counts. - 3Third-party software represented 77.4% of observed AI packages and tools, making dependency and connector visibility part of AI governance. ## Scoring Rationale The report provides a large telemetry-based view of enterprise agents, MCP infrastructure and AI dependencies, with direct relevance to asset discovery and software-supply-chain security. It is useful to platform and security teams but does not introduce a new model or product capability. ## Sources Primary source and supporting public references used for this report. Practice interview problems based on real data 1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with. [Try 250 free problems](/problems)