Architecture AI Updates: August 25, 2026

1. A Capability-Based Security Model for Zero-Trust AI Agents

InfoQ / Cloudflare. Writing on InfoQ, Bruno Couriol details Cloudflare’s open-source internal AI platform, which replaces the shared-SaaS model with per-user isolated application instances running in dedicated V8 isolates on the workerd runtime. Its “Gatekeepers” layer applies a capability-based security model where agents begin in a zero-trust state with zero ambient permissions, then gain narrowly scoped access through column masking, role-based rate limits, and human approval for destructive operations. The design differs from typical Model Context Protocol setups by ensuring shared gadgets can only touch resources the recipient already holds, preventing accidental privilege leakage across tenants. Source

2. Treating AI Governance as a Runtime Loop Instead of a Policy Document

InfoQ / Microsoft. On InfoQ, Leela Kumili describes Microsoft’s shift from static AI policy documents toward runtime enforcement and continuous monitoring across nine governance domains including data, model, agent, observability, and identity. The architecture frames governance as a continuous operational loop where policies set requirements, controls enforce them at runtime, observability captures behavior, and evaluations test quality before and after deployment. Microsoft Foundry’s AI Gateway acts as the enforcement boundary, applying authentication, quotas, and policy without requiring changes to application code. Source