Mintive guide to non-human identity management for AI agents: types, controls, lifecycle

In two minutes

• The guide argues that traditional human-centric IAM does not fit AI agents, which authenticate at machine speed, may spawn sub-agents and have no HR trigger for their lifecycle.
• It says agents should be treated as first-class principals with their own identities, credentials, named human owners and audit trails.
• It describes five agent identity types from the Cloud Security Alliance's draft framework (copilot, autonomous, orchestrator, ephemeral sub-agent, agent-as-a-service), each with different credential needs.
• It recommends a permissions-first approach: deny by default, unique identity per agent, task-scoped permissions, a tool allowlist, and just-in-time, time-bound access that revokes after the task.
• It recommends zero trust checks on every tool call, screening of tool outputs for credentials, and policy enforcement outside agent code.
• It sets out a lifecycle: discovery, provisioning, ownership, credential management, review and decommissioning, and lists the evidence auditors will ask for.
• Much of the text describes Mintive products (Agent Gateway, MCP Gateway, Agent Monitor, Mint Guard, Gateway Middleware) as the way to apply these controls.

Key numbers

• more than 90 to 1: ratio of non-human identities to human identities in many organizations (section Key Takeaways).
• 91%: organizations with limited or no visibility into their AI agent identity populations (section Visibility and Auditability).
• 50%: organizations that have already had a security incident related to non-human identities (section Understanding Non-Human Identity Management for AI Agents).
• 78%: organizations lacking formal policies for creating or removing AI identities (section Understanding Non-Human Identity Management for AI Agents).
• 28%: organizations that can trace agent actions back to a human sponsor when incidents occur (section Understanding Non-Human Identity Management for AI Agents).
• December 2, 2027: date EU AI Act Annex III high-risk requirements apply (section Visibility and Auditability).
• August 2, 2028: date requirements apply for high-risk systems embedded in regulated products under the EU AI Act (section Visibility and Auditability).

What to do with it

• Inventory agents first, starting with those holding the highest privilege, touching sensitive data or running without human oversight, and treat orchestrators as the top priority (FAQ).
• Assign each agent a unique identity and a named human owner, and stop agents running on shared or borrowed human API keys (least privilege and credential sections).
• Match authentication to agent type: session-bound for copilots, short-lived for autonomous agents, mutual TLS or SPIFFE SVIDs for orchestrators (agent types and authentication table).
• Move to just-in-time, time-limited grants with automatic revocation, and cap delegation depth so sub-agents cannot exceed the parent's privileges (credential lifecycle and FAQ).
• Screen tool outputs for embedded credentials and require re-authorization before an agent uses a discovered credential (runtime guardrails section).
• Build audit evidence now: provisioning records, access timelines, certification history, delegation chains and decommissioning proof (FAQ on audits).

Read the full document if

• you need a checklist of controls for governing AI agents as non-human identities.
• you want the five-type agent identity taxonomy and the authentication options mapped to it.
• you are comparing what an agent gateway product covers, such as VMCPs, guardrails and SIEM export.

What it relies on

The guide gives no method, sample or period for its statistics, and it cites no source for the 90 to 1, 91%, 50%, 78% or 28% figures. It names outside references: the Cloud Security Alliance's draft Agent Identity Governance Framework, the OWASP Agentic Security Initiative (ASI03), the EU AI Act, DORA and the NIST AI Risk Framework. The publisher, Mintive (formerly MintMCP), sells agent gateway, monitoring and guardrail products of the kind the guide recommends, and the guide presents them as the way to apply its advice.

Our read

Useful as a control checklist and a vocabulary for agent identity types, delegation and credential lifecycle. The statistics have no stated source, so check them before quoting. Product sections show one vendor's design and are not a neutral comparison.