
November 23–27, 2026
Cloud Developer
Psquare
Microsoft AI Foundry's Toolbox makes connecting MCP tools to an agent trivially fast a connection and a version bump, no code changes. But nothing in that flow stops a mutating tool from shipping without a human-approval gate, a connection from using a static shared credential instead of an identity-backed one, or a sample tool output from quietly leaking a real customer's PII into your agent's context. In this session, I'll walk through foundry-toolbox-radar-lab, an open-source CLI I built that scans a Toolbox configuration for exactly these risks ungated mutating actions, static credentials, missing Entra audiences, PII/secret leakage, and overly broad scopes before deployment, with a severity-tagged report and a CI gate that blocks the merge automatically. You'll leave with a practical mental model for governing agentic tool access in Foundry, a free tool you can run against your own Toolbox configs today, and a working GitHub Actions pattern to turn that check into an automatic gate. Live demo included I'll run the scanner against a deliberately under-governed config and show exactly what it catches.
See every session across the AI Agents Conference November 2026 line-up.
View Agenda