Approved Isn’t Safe: What 3,000+ Enterprise Codebases Reveal About Model Exposure and Agent Reach
Date: Tuesday, December 1 at 8 am PT / 11 am ET
Snyk matched AI-BOM scans from 3,023 enterprise environments against adversarial test results for named models. In 74.5% of the 1,693 environments that reference a model, at least one model scores high. About one in four model occurrences lands in the high band, whether the estate holds 3 models or 300. The exposure stays flat. What grows is what those models can reach.
In this session, Snyk’s John Carione walks through Volume III of the State of Agentic AI Adoption research. It isn’t a product demo. You’ll leave knowing which models on your approved list to review first, when swapping a model reduces risk, and when only controls do.
In this session, you’ll learn:
- Why your approved list is the first place to look. The older, lower-capability models in the “conservative core” make up about a third of scored occurrences. They score worse on personal data leakage, decision override and insecure code than the frontier models most teams worry about.
- When switching models fixes the problem, and when it doesn’t. On command execution, 77% of tested models score zero, so a swap helps. On misinformation and race bias, results cluster near the ceiling. There, access limits, output validation and human approval do the work.
- What provider, license and price tell you. Among mid-range models, provider averages run from 113 to 356. At the frontier, three leading providers land within 22 points of each other. Open-weight models score about twice as high on personal data leakage (562 vs. 274). Price predicts nothing.
- The difference between exposure and reach. The score is a property of the model. Reach comes from agents and MCP servers. Of 1,051 US accounts with an MCP reference, 744 build a server, not just connect to one.
The supply chain your tests can’t see. 904 of 1,693 model-bearing environments can’t say what data trained their models. Open-weight lineage is a dependency to verify. For proprietary models, it’s a question for procurement.
Register now!
About the presenters:
John Carione, Senior Product Marketing Director, SNYK
Date: 12/01/2026
Time: 8:00 AM PT