Products & Tools

Tanium launches AgentSight to inventory shadow AI on employee devices

Tanium’s AgentSight now inventories local models, coding agents, extensions and MCP servers across three desktop platforms, but behavior analysis and enforcement remain roadmap items.

Tanium AgentSight Brings Shadow AI Into the Endpoint Inventory
Tanium AgentSight Brings Shadow AI Into the Endpoint InventoryAI-generated
By James Calloway6 min read

Why it matters

  • AgentSight’s first capabilities are available now in the Tanium console.
  • The inventory covers four categories across Windows, macOS and Linux.
  • Each discovered item is associated with a device and the person who installed it.
  • Initial visibility supports on-premises and air-gapped environments.
  • Behavior analysis, session-and-prompt attribution and policy enforcement remain on the roadmap.

By James Calloway

Tanium has released the first AgentSight capabilities in its console, which can identify four categories of AI software across Windows, macOS and Linux. Tanium says the features are “available now in the Tanium console.”

Tanium introduced the product at its Converge conference in Las Vegas this week. The launch targets a gap that centralized purchasing and cloud records leave open: employees can install AI tools on individual workstations without those tools appearing in a company-wide subscription list.

That gap matters because installation and authorization are different questions. Tanium says AgentSight can show what is present and where it resides, but it cannot yet confirm that an agent’s actions are authorized. Its first release supplies visibility, not proof of compliance.

What can AgentSight inventory today?

Tanium says the initial inventory covers four categories:

  • Local models running on the employee device.
  • Coding agents used for software-development tasks.
  • AI extensions added to other applications.
  • Model Context Protocol (MCP) servers, which connect AI tools to other services.

Tanium associates each item with its device and the person who installed it. The company calls this endpoint inventory: a catalog of AI software across employee devices. The result can show security teams where a tool lives and whom to contact about it.

Tanium also says cost tracking in Tanium Asset shows where AI software is deployed and what it costs. That gives procurement teams a deployment and spending view, but it does not establish whether the tool is approved. Nor does it explain what an agent did or whether its activity followed company rules.

The announced details provide no performance benchmark or coverage percentage for discovery. The initial scope also stops short of linking an agent’s behavior to an authorized session. Presence on a device is not a governance decision.

Takeaway: AgentSight maps AI software across three operating systems, but the available evidence supports an inventory claim rather than complete control.

Why does device-level visibility matter?

An enterprise can govern a company-wide AI subscription and still have limited visibility into assistants staff install on their own workstations. An approved coding tool might sit beside several extensions, a local model and a connector to another service.

A software inventory lets IT see which devices run those components and provides a starting point for tracking down who installed them. Approving one service does not automatically account for every tool that enters a workflow, particularly when employees add extensions or local components independently.

Tanium says the initial visibility supports on-premises and air-gapped environments. Air-gapped systems cannot send telemetry, meaning device data reported for monitoring, to a cloud service. That constraint matters to teams operating without continuous cloud connectivity.

The endpoint view does not replace records from browser-based services, cloud systems or applications an agent uses remotely. It therefore complements those records rather than superseding them, leaving administrators with several incomplete views that still require reconciliation.

Takeaway: Endpoint records close a local visibility gap, but companies still need cloud, browser and remote-application records for a complete picture.

What did the Converge demonstration establish?

Tanium described a Converge demonstration of a “Shadow AI” page in Atlas. In that demo, security teams could see changes in the inventory and ask questions about them in plain language. The interface presents device-level findings without requiring administrators to inspect every system manually.

Tanium also described plans to rank findings and provide explanations. It plans to tie agent activity to the originating session and prompt, which would let investigators trace behavior farther back than a software name. Those features would make unusual activity easier to investigate and policy exceptions easier to explain.

The company said the first capabilities are available, but the demonstration and roadmap do not prove those additions are generally available. Nor do they establish coverage for every agent or service. A working interface shown at a conference is evidence of intended direction, not validation of enterprise-wide performance.

The launch details provide no general-availability date for the planned behavior features. That leaves buyers comparing a live inventory system with governance functions that remain under development, rather than two capabilities operating at the same maturity level.

Takeaway: Converge showed a usable inventory interface, while most of the promised investigation workflow remains a roadmap.

When can Tanium enforce agent policies?

Tanium’s long-term product is intended to explain what an agent did, flag unusual behavior and apply policy across many devices. The company is still developing those parts of the system, and none belongs to the initial release described at launch.

Tanium plans to let security teams define policies centrally and detect violations before turning on enforcement. That sequence gives administrators time to test detections before they can block legitimate work. It also makes false positives a product issue rather than merely a policy-writing mistake.

A local tool that looks unusual in the inventory may support a legitimate engineering or accessibility workflow. Reviewing those cases with the people who own the affected systems can help organizations distinguish authorized exceptions from unmanaged software without disrupting routine work.

None of that changes the current evidentiary limit. AgentSight can show that a component exists and associate it with a device and installer, but it cannot yet establish that an agent acted within policy. An inventory identifies candidates for review, not violations.

Takeaway: Policy enforcement is planned, but today’s inventory cannot tell approved activity from unauthorized behavior.

How can an inventory become governance?\n

The voluntary NIST AI Risk Management Framework, a framework that evaluates AI risk through system use, puts risk in the context of how systems are used rather than simply what is installed. For enterprises, that means linking endpoint inventory to the people, permissions and business processes behind each entry.

Adoption also raises a data-governance question about what information the system should collect and who may review it. If future activity records include prompts, those prompts may contain confidential work and require deliberate access controls. Inventory records also need retention and review rules that fit the organization.

The near-term test for AgentSight is whether security teams can maintain an accurate inventory as employees adopt and change tools. AgentSight must recognize new components, retain the device and installer association and avoid presenting stale entries as current facts.

The more demanding test is whether businesses can turn that visibility into useful decisions. Security leaders must determine which tools are approved, where an exception is justified and who should review behavior outside policy. Product support alone will not resolve those organizational questions.

Tanium has shipped the first part of this work. The rest will depend on both the product roadmap and how companies put the inventory to use. Any claim of comprehensive AI governance would run ahead of both the available features and the operational work required to use them responsibly.

Takeaway: AgentSight can support governance, but only after companies add permissions, human review and trustworthy activity records.

Source: unite.ai

Share this article:

More from James Calloway

James Calloway

Show full bio

News editor covering people, launches and day-to-day practice in artificial intelligence and machine learning at AI In Context.

791 articles

Related articles

  1. Tanium launches AgentSight to track and govern AI tools on endpoints
  2. Tanium launches AgentSight to track AI agents on corporate devices
  3. Virtana’s AI agent war room targets cross-system IT incidents
  4. Partnership on AI: Agent Frameworks Can't Show What Agents Did
  5. AI Agents Move Onto Windows PCs, and Permissions Become the New Battleground

« Previous articleNext article »