What happened
On Nov 10, 2026 Obsidian Security announced its Runtime Policy Decision Engine, which evaluates an AI agent's proposed action against centrally managed policies in the instant before it executes, using context about who is behind the agent, its entitlements, and the third-party app / data classification involved. It ships an MCP Security Proxy that intercepts agent tool calls so app tokens never reach the AI client, a Policy Decision API that extends Obsidian's verdict to customer-operated gateways, runtime policy support for OpenAI Codex (joining Claude Code/Cowork and Microsoft Copilot), and discovery/governance for in-app agents inside Salesforce Agentforce, Databricks, GitHub, M365, Snowflake, ServiceNow and Workday.
Why it matters
This is one of the first context-aware, decision-vs-enforcement-split runtime engines that stops an otherwise-authorized agent from taking an over-scoped or destructive action — the exact gap simple access control can't close. It also standardizes MCP tool-call interception (with Cloudflare's MCP Server Portal) so enterprises can govern agent traffic without a separate rule set.
Applicability
Enterprises running coding agents (Codex/Claude Code/Copilot) or SaaS-native agents should evaluate Obsidian for runtime enforcement and MCP token brokering now; useful for anyone already doing agent discovery and needing to move from visibility to allow/block.