Safety & Security

Tanium launches AgentSight to track AI agents on corporate devices

Tanium's AgentSight inventories AI agents, models and MCP servers on Windows, macOS and Linux endpoints, links them to users, and logs session-level actions with tamper-evident records.

Tanium launches AgentSight to track AI agents on devices
Tanium launches AgentSight to track AI agents on devicesmwitt1337 / pixabay
By James Calloway5 min read

Why it matters

  • Tanium launched AgentSight on 18 November 2026 to track AI agents, models and MCP servers on Windows, macOS and Linux endpoints.
  • The product works through the Tanium console and supports on-premises and air-gapped environments.
  • AgentSight records what each agent did in order during a session, plus the prompt that started it, in a tamper-evident log.
  • Policy controls will ship with a detect-only mode before any automated enforcement is applied.
  • Cost tracking for AI software is added through Tanium's Asset product for IT and finance teams.

Tanium has launched AgentSight, a product that gives organisations visibility into AI agents, models and MCP servers (Model Context Protocol servers — local connectors that let AI tools interact with other software and data) running on endpoint devices across Windows, macOS and Linux.

The launch, announced 18 November 2026, targets a problem IT and security teams already face: employees install AI tools on work devices without central oversight. Those tools include coding assistants, local models, desktop AI applications, browser extensions, IDE extensions and MCP servers. AgentSight extends Tanium's endpoint monitoring discipline to this category of software.

The stakes are straightforward. Unlike standard applications, AI agents may act using the permissions of individual employees, which makes it harder for security teams to distinguish normal activity from misuse. Tanium's bet is that endpoint-level monitoring stays central to solving that problem, because many AI tools run locally on employee devices rather than only through centrally managed cloud services.

What does AgentSight actually do?

Available through the Tanium console, AgentSight identifies AI components on devices, links them to the machines where they run, and associates them with the person who installed them. It supports on-premises and air-gapped environments — a deliberate pitch to sectors where data handling restrictions or operational requirements limit the use of internet-connected services.

Tanium positions the product as a way to manage AI use on endpoints with the same approach organisations already apply to devices. That means visibility into what is installed, records of activity, and risk findings tied to individual agents.

The product creates an inventory of AI software discovered across a device fleet and highlights day-to-day changes. It also adds cost tracking for AI software through Tanium's Asset product, giving IT and finance teams a view of spending and deployment across the organisation.

The takeaway: AgentSight is an inventory and audit layer first, not a standalone security tool.

How does it spot a misbehaving agent?

Harman Kaur, Chief Technology Officer at Tanium, said the company is applying its existing behavioural monitoring methods to AI software operating on endpoints.

"An agent misusing its access may look like legitimate activity, so you need to know what's expected and normal. Tanium already builds behavioral baselines for endpoints, and AgentSight extends that discipline to agents," Kaur said.

The core assumption is that anomaly detection beats rule-based blocking: an agent acting with a user's permissions does not automatically look malicious, so the system has to know what normal looks like for that specific agent. That claim rests on the quality of the baseline — Tanium has not yet published detail on how quickly baselines form or how they behave with rarely used agents.

Kaur said the system ranks AI risk findings, explains them, and keeps a record of what AI tools did during a session.

"It will rank AI risk findings and give a plain reason for each. It will also record what each agent did, in order, in the session and the prompt that started it, and alert when an agent departs from its own normal. Endpoints, agents, the infrastructure they depend on, and the identities behind them are managed as one estate on one platform," Kaur said.

In plain words: the pitch is a session-level audit trail — action, order, initiating prompt, and the human behind it — rather than a simple allow/deny decision.

What enforcement powers come with it?

Tanium is building policy controls for AI software on endpoints. These let administrators define a policy once and then identify devices that fall outside it. Notably, the framework will include a detect-only mode before any automated enforcement is applied — an acknowledgement that blocking employee AI tooling outright can break workflows before teams understand their own usage.

Some enforcement and response tools already exist elsewhere in the Tanium platform:

  • Process termination
  • Device quarantine
  • Live Response functions through Threat Response, for organisations that need to act on a device quickly

So the enforcement story spans products. AgentSight supplies the visibility; existing Tanium modules supply the muscle.

Where does it sit in Tanium's portfolio?

AgentSight is a separate product alongside Tanium's Security Operations, Endpoint Management, Exposure Management and Atlas offerings. Atlas is the company's AI layer and can query findings in natural language and explain results shown on a Shadow AI page — Tanium's own term for unmanaged AI use surfacing inside an estate.

The naming of a dedicated "Shadow AI" page is itself a signal: Tanium treats unmanaged AI as a distinct visibility problem, not a subset of generic software inventory.

Can the records be trusted?

Tanium said it keeps a tamper-evident record of what each agent did and who it acted for. The company frames this as an extension of the data it already tracks for endpoints, treating agents as objects with an identity, state, owner and history.

That framing matters for incident response and compliance: if an agent took an action with a user's permissions, the record needs to show both the machine, the identity, and the sequence of actions — not just a log entry. Tanium has not detailed the technical mechanism behind the tamper-evidence claim.

Why now?

The launch reflects a wider shift in enterprise technology management: companies are trying to bring AI use under the same controls applied to conventional software and hardware. Employees adopting AI tools faster than policy can follow is the driver, and vendors across endpoint management are converging on similar answers.

Tanium's differentiators are its endpoint-native approach — tying agents and models back to individual devices and users — and support for environments that cannot rely on cloud connectivity. Whether behavioural baselines built for endpoints translate cleanly to agent behaviour is the open question; Tanium's answer is that the discipline extends, but early deployments will test that premise against agents whose usage patterns change weekly.

The detect-only mode suggests Tanium expects customers to spend a discovery period mapping their AI estate before turning on enforcement — which is likely the honest sequence for most organisations.

Source: itbrief.com.au

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.

722 articles

Related articles

  1. Tanium launches AgentSight to track and govern AI tools on endpoints
  2. AI Agents Move Onto Windows PCs, and Permissions Become the New Battleground
  3. Virtana’s AI agent war room targets cross-system IT incidents
  4. OpenAI Launches Workspace Agents in ChatGPT for Teams
  5. Airlock Digital adds action-level controls for Codex and Claude

« Previous articleNext article »