Verify the agent, secure the device: a closer look at governing AI end to end with Jamf and Okta 

AI agents don't live in one place. A Claude Code session on a developer's Mac authenticates to an identity provider, calls tools over MCP, reads files on disk and sends data to a model endpoint. Each step has its own control point. At Oktane and JNUC 2026, Okta and Jamf each announced controls for the half they own.

Okta covers the identity and runtime path of the agent. Jamf covers the device the agent runs on. Here is what shipped, what's coming, and how the two halves reinforce each other.

The problem: agent activity spans two control planes

Jamf's AI Governance Survey found that organizations with deeply integrated AI are 40% more likely to report an incident than those still exploring it. That tracks with what admins see in practice. AI tools install locally, store config in user space, authenticate with personal accounts, and reach corporate data through browsers and tool connections that traditional controls were never built to inspect.

Two questions come up in every governance review:

1. Who or what is this agent, and what can it access at runtime? Okta for AI Agents (Identity & Access)
2. Where is it running, and is that environment managed and compliant? Jamf (Device)

Answering only one leaves a gap. A valid agent identity on an unmanaged device is a risk. A fully managed device tells you nothing about what an agent should be allowed to reach.

Okta: agents become first-class identities with runtime authorization enforcement

At Oktane, Okta extended Okta for AI Agents along 4 pillars:

Where are my agents? (Development, Discovery & Identity)

Okta already discovers shadow AI agents and their connections across managed browsers. Shadow AI Agent Discovery for Endpoints extends discovery to devices where unmanaged agents run without IT and Security’s knowledge so you can bring them under your control.

What can they do? (Access and Entitlements)

Agent Single Sign-On (SSO) makes the open Cross App Access standard part of core Okta SSO, giving agents a managed sign-in path, so connections to downstream apps are established through the identity provider rather than through ad hoc, user-granted OAuth consent. Agent-to-Agent Connections sets rules for which agents can call other agents, what each agent can access, and captures the handoff in an auditable chain. Configuration Designer visually maps agent-to-resource connections so the handoff is governed, scoped, and auditable. Resource Access Certifications reviews agent connections over time to prevent standing and excessive permissions.

What are they doing? (Runtime Authorization & Control)

Agent Gateway sits in the execution path between agent and tool call, enforcing policy and logging every interaction at runtime.

How do I respond? (Active Containment)

Admins can already deactivate an agent in the admin console, which blocks new sessions. Okta will also expand kill switch capabilities to Okta’s Agent Gateway. Deactivating an agent there revokes every active token it holds and terminates sessions in flight. That distinction matters technically: blocking new sessions does nothing about a bearer token that is already issued and valid. You need to make sure that every new request is blocked.

Okta also announced the Blueprint Alliance, an industry group working toward shared principles for securing agents. It gives security architects a common reference as the standards take shape with a commitment to interoperability across the agentic stack.

Jamf: AI governance at the device layer

Jamf AI Governance has been available in Jamf for Mac since June. It works through the management channel admins already use. Jamf monitors supported AI platforms for new or changed controls and translates governance intent into vendor-correct configuration on macOS. Initial support covers Claude Code, Claude Desktop and OpenAI Codex. The result is three capabilities:

Visibility

Discover which AI tools are running on managed Mac devices.

Control

Set and enforce policy for how those tools can be configured and used.

Reporting

Generate audit-ready output for internal governance, boards and regulators, including activity insights and cost and model insights.

At JNUC, Jamf extended the model:

Browser

Visibility and control across Safari, Chrome, Edge, Firefox and other browsers, aimed at detecting personal AI accounts and enforcing policy at the point of action, such as stopping a corporate file from uploading to a personal AI account. Jamf has not announced a date.

Mobile

Detect, monitor and block AI use in native apps and browsers on iPhone and iPad devices, built on on-device content filtering and device management. General availability is targeted for Q1 2027.

The platform underneath: programmable by design

Governance only scales if you can automate it. The JNUC keynote framed Jamf as a platform you build on, and the changes are concrete:

Platform API Gateway

One endpoint and one authentication model for JamfAPIs, with credentials managed centrally in JamfAccount. It covers the blueprints API, the compliance benchmark API and device APIs. It also works as a migration path off the Classic API.

Jamf Platform Terraform Provider

Manage a Jamf instance as code. Lloyds Banking Group already runs it in production, which makes AI governance policy something you can version, review in a pull request and roll back.

Routines

Moving to a Jamf-hosted automation engine with more than 1,500 integrations, better visibility into automations and control over data residency. Device health. Continuous assessment of memory, battery and crashes to flag at-risk Mac devices and tie findings to remediation.

Ring Deployments

Automated Ring Deployments let organizations test updates with a pilot group, pause a rollout when issues are detected, and expand to the broader fleet only after admin approval. Public beta is targeted for Q1 2027.

Better together: a closed loop from device to agent

Put the two together and you get a control loop instead of two separate dashboards.

1. Jamf establishes and maintains device trust. The device is enrolled, managed and compliant, and sanctioned AI tools are configured to policy and unsanctioned AI tools are surfaced for appropriate action.
2. Okta makes access decisions with that context. Agent identity, user identity and device posture inform whether an agent's request is allowed, evaluated at sign-in through Agent SSO and at runtime through Agent Gateway.
3. Enforcement happens on both sides. If an agent misbehaves, Okta can revoke its tokens and end its sessions. If a device falls out of compliance, Jamf can remediate it, and that change in posture can feed back into access decisions.
4. Everything lands in a shared audit trail. Jamf reports what ran on the device and under what policy. Okta logs what the agent accessed and when.

Jamf's June announcement described the intent directly: anchoring Okta for AI Agents to Jamf's endpoint enforcement so that each agentic connection on a managed Mac is authenticated, authorized and visible from the device to the data.

Neither vendor can see the whole path alone. Okta establishes identity with a view into which downstream resources an authenticated agent is calling, while Jamf provides insight into a local config file or a browser upload. Together, they cover both.

What to do next

1. Inventory: If you use Jamf for Mac, run AI Governance visibility and see what's actually on your fleet.
2. Audit how agents authenticate today: List where AI tools hold long-lived tokens or user-granted OAuth consent, and evaluate Agent SSO as the replacement.
3. Define policy once, enforce twice: Write down which AI tools are approved, on which device states, with access to which resources. That policy needs a device-side and an identity-side expression.
4. Move toward the Platform API: Start mapping Classic API usage to the Platform API Gateway and consider managing governance configuration through Terraform.
5. Track the dates: The Okta gateway kill switch and Configuration Designer land in Q4, and AI Governance for Mobile is likely Q1 2027.