Skip to content
IAM Insights

Explainers · Glossary

MI9

A runtime governance framework for AI agents, from a 2025 arXiv paper, that keeps watching an agent after approval through telemetry, authorization monitoring, drift detection and graduated containment.

reference →

In our briefings

The claim each of the newest briefings on MI9 would have you leave with.

  1. Ibrahim and Li argue that tokens are the wrong place for agent delegation, since a token's contents are fixed when it is issued, and instead add agent, session and scope types to an existing relationship-based permission model without deleting anything in it. An agent's permission on a resource becomes what a delegating human holds there, intersected with the scope of the agent's active session. Their Theorem 5.3 proves soundness: any permission an agent gets this way implies a human with that permission, a valid delegation path from that human, and a scoped session. It does not prove that authority narrows at each hand-off. The paper defines attenuation, leaves its ordering to each domain, and lists a formal treatment as future work.

    Huawei researchers add agent delegation to an existing permission model and prove each action traces to a humanBI-FIND-2026-0897

  2. Kadaboina argues that OAuth and mutual TLS settle who may call an agent's capability and say nothing about the terms of use, which change as the service revises its policy. In his reading, "Neither A2A nor MCP defines a versioned usage policy object, a typed consent record, or a mechanism to block skill invocation until a new policy has been processed." His proposal adds three records: a versioned policy, a consent record that must hold an entry for every clause, and a per-action log of which clause the calling agent checked and what it decided. A service is to refuse a skill when the caller has disputed, or not understood, a clause that governs it.

    A proposed consent log for A2A and MCP records the terms an agent accepted, but the agent fills it in itselfBI-FIND-2026-0975

  3. MI9's authorization layer re-evaluates an agent's permissions whenever its goal or risk level changes. The authors list four controls: authority matrices saying who may delegate which permissions, delegation chains checked for provenance and given an expiry, ordering guards such as approval before use, and default-deny for a goal change that would raise the agent's risk tier. When a parent agent's tools are restricted, they write, revocation may have to cascade to the subagents it spawned. What sits underneath is thin: the agent's identity is a metadata field on telemetry events, the paper never mentions a credential or an authentication step, and the "formal security guarantees" it claims for the delegation graph are stated, not proved.

    MI9 re-checks agent permissions when goals change; its 99.81% score comes from tests Gemini wrote and scoredBI-FIND-2025-0140

Cited in 3 briefings

Every briefing that discusses this term, newest first.

  1. Huawei researchers add agent delegation to an existing permission model and prove each action traces to a human

    Paper · · arXiv · Amjad Ibrahim et al. · Jun 2, 2026

  2. A proposed consent log for A2A and MCP records the terms an agent accepted, but the agent fills it in itself

    Paper · · arXiv · Ravi Kiran Kadaboina · Apr 16, 2026

  3. MI9 re-checks agent permissions when goals change; its 99.81% score comes from tests Gemini wrote and scored

    Paper · · arXiv · Charles L. Wang et al. · Aug 5, 2025

Related terms

What appears alongside MI9 in the same briefings.

IBACLangGraphAutoGen