Tanium has introduced AgentSight, a product intended to let corporate IT teams identify and control artificial intelligence agents, local models and Model Context Protocol servers running across employees’ devices. The 17 November announcement addresses a problem emerging as AI tools become part of everyday work: an organisation may struggle to know what software is making decisions or accessing information on its computers.
The concern is sometimes called “shadow AI”, a variation of the long-standing problem of employees adopting technology without formal IT approval. A browser extension that summarises files, a coding assistant with repository access or an agent connected to business applications can all create security questions, even when employees use them to perform legitimate work.
Why AI agents complicate device security
Conventional endpoint security aims to identify suspicious files, processes and behaviour on computers and servers. AI agents add another layer because they may receive instructions in natural language, connect to external services and act on behalf of a user. The device may appear normal while a connected tool has broader data access than security administrators intended.
Model Context Protocol, or MCP, has become one way of connecting assistants to tools. An MCP server can expose a capability such as reading a repository or fetching a business document. That can be useful, but businesses need to know which servers are installed, what they can access and how their permissions are controlled.
Tanium says AgentSight extends its existing endpoint visibility and management capabilities to AI-related components. The company describes an ability to discover agents, models and MCP servers, observe their activity and apply governance controls. As with any security tool, the value depends on coverage, accuracy and how reliably policies can be enforced in a customer’s actual environment.
What discovery can and cannot solve
Finding an unauthorised agent is a necessary first step, but discovery alone does not establish that the agent is malicious. An employee may install a useful tool that lacks formal approval, while a permitted application can still mishandle information. Security teams therefore need risk-based decisions rather than a blanket assumption that all AI tools are dangerous.
Administrators should also consider whether controls can distinguish authorised from unauthorised behaviour, capture meaningful audit records and restrict sensitive operations. Strong employee authentication remains important because an AI tool acting under a legitimate account may inherit the privileges associated with that account.
There are privacy and organisational questions too. Device-level monitoring can create employee concerns if collection practices are unclear. Companies should define appropriate purposes, access rights and retention periods while complying with applicable law and workplace policies.
Enterprise implications in Asia
In Singapore and across ASEAN, companies often combine central IT systems with regional teams using different software. That makes visibility difficult, particularly as staff adopt new AI products faster than security policies are updated. A common inventory of AI tools could help organisations decide what to permit, what to restrict and which integrations need additional safeguards.
However, the company has not supplied independent comparative measurements showing that AgentSight will eliminate a defined proportion of incidents. Buyers should ask which operating systems are supported, how discovery works without disrupting work, what types of AI calls are visible and whether controls apply equally to local and cloud-based agents.
The announcement illustrates a shift in cybersecurity priorities. Defending endpoints is no longer only about preventing malicious code from running. Organisations also need to govern legitimate software that may be powerful enough to act unexpectedly. Tools such as AgentSight may contribute to that oversight, but training, clear policy and careful permission design remain essential.
Sources and further reading
Sources: Tanium product announcement, 17 November 2026; Business Wire distribution, 17 November.