Your agents have standing access to everything. The fix is to give them almost nothing, briefly.

The access problem

AI agents currently operate on standing, over-scoped access that does not require an adversary to exploit, according to Cyera's account. The current OAuth consent model is designed for human users and lacks primitives for task-based, machine-speed, non-reviewed access. A human clicks approve once and understands the scope; an agent acts thousands of times a minute on a token granted for something broader. The risk is not that someone hacks the agent. It is that the agent does exactly what its token allows, which is too much. m:…1392

Cyera proposes moving to a centralized, broker-based model using Enterprise Managed Authorization, known as EMA, and Identity Assertions JWT Authorization Grant, known as ID-JAG, to issue task-specific, short-lived credentials. EMA became a stable MCP extension on June 18, 2026, and ID-JAG is an active IETF OAuth Working Group draft currently at revision 06. Token Exchange, RFC 8693, is the mechanism that enables per-task scoping at boundaries, and Cyera Identity evaluates agent intent against policy before any token is requested. The OpenID Foundation is running an MCP Security Interoperability event with Gartner to test EMA and ID-JAG implementations. The catch is adoption: support for EMA and ID-JAG is currently uneven across identity providers and MCP server vendors, and enterprises should maintain one identity control plane even if enforcement points are plural. m:…1392

What it looks like inside a product

Harvey, the legal AI platform, shows what this kind of enforcement looks like when it is built into the product rather than bolted on. It enforces ethical walls, information barriers, on conversations, tables and drafts created from documents within its platform. It separates matter identity sync from access policy sync to allow independent scheduling, and uses a centralized policy engine to ensure consistent wall enforcement across all product surfaces. Its largest workspace contains approximately 1.7 million matters, and it uses a dedicated search index built on turbopuffer to power the matter picker. The index contains searchable text but does not store wall data, and if a wall fetch fails for a matter, Harvey retains the last stored policy rather than failing open. m:…8405

Harvey also encrypts customer data at rest using AES-256 and is contractually prohibited from training AI models on customer data. Those are table stakes, but the wall enforcement is the interesting part: it is a system that treats access policy as data that must be synced, indexed and enforced consistently, not as a setting in a menu. Airia's account of Microsoft's Agent 365 shows the gap this fills. Agent 365 became generally available for commercial customers on May 1, 2026, and uses Entra for identity and risk-based access controls, but Microsoft Purview's classic DSPM for AI does not support Agent 365, and content created by Agent 365 agents does not inherit sensitivity labels from source files. Agent 365 works best with Microsoft E5 as a prerequisite. The native controls cover the Microsoft stack; the gaps are everywhere else. m:…8405 m:…5942

Governance is a loop, not a gate

Airia's proposed answer is to extend Microsoft's controls outward: it acts as an external threat detection provider for Copilot Studio, enabling policy enforcement on tool calls, integrates with Microsoft Sentinel to stream agent activity and policy violations, and provides a routing engine for model access with budget controls. It can run as shared SaaS or as a private deployment in an Azure tenant, and is transactable on Azure Marketplace, counting toward Microsoft spend commitments. m:…5942

LangWatch adds the last piece: governance has to learn. Its data flywheel is a loop where using the product produces the data that improves it, in five steps: trace, evaluate, turn the flag into a test, fix, and rerun on every change. Scenario tests are required for failures that depend on previous conversation turns, and CI integration can be configured to exit non-zero and fail the build when a scenario fails. Predicting all edge cases for an LLM product beforehand is impossible, according to its account, so the discipline is to turn every failure caught in production into a test. Access control decides what an agent may do; the flywheel decides what it will do better next time. Together they are the shape of enterprise agent governance this week: narrow credentials, enforced policy, and a loop that learns from every failure. m:…0519