Opinion

AI Is Moving Onto Our PCs — But Who Controls What It Can See?

Adlin Pertishya
Add Unite.AI to your preferred sources on Google
Key PointsLocal AI processing does not automatically guarantee privacy or security for on-device…
View more
  • Local AI processing does not automatically guarantee privacy or security for on-device agents.
  • Agents must operate under least-privilege access limited to what each specific workflow requires.
  • Users need transparent and revocable control over what agents can see and do.

AI agents are beginning to move beyond chat windows and into the operating systems, applications and files people use every day.

New Windows PCs and software released in recent months are ushering in a new era of multistep autonomous AI workflows for the ordinary user and into enterprise environments, and the transition is no longer just a technology forecast.

NVIDIA’s RTX-powered Windows PCs support locally running AI agents, while Perplexity has expanded its agent capabilities to compatible Windows computers, enabling workflows involving local files and connected applications. Meanwhile, Microsoft is developing safeguards to control how agents access resources and take action.

But greater autonomy brings greater security risks. On October 19, Microsoft’s Executive Vice President of Security, Hayete Gallot, warned that organizations must do more to secure “agentic insiders,” noting that AI agents are frequently overprivileged and can access sensitive data, tools and systems while making decisions in critical processes.

Microsoft’s Windows security strategy is based on containment, limited permissions, limited oversight, and increasing concern that an agent’s error could have repercussions beyond a simple wrong answer. It may reveal sensitive information, alter critical files or cause unwanted activities within business applications.

Local AI Does Not Automatically Mean Private AI

Using AI agents on a PC can help minimize the amount of sensitive data being transferred to cloud-based services, but local processing is not an all-out security safety net.

Even an on-device model can still have too many files accessed, poor security or unsafe links to other apps.

The problem isn’t always the AI model; it’s often the organizational governance, said Bayron Quintero, CEO and co-founder of iData Global, a company specializing in AI solutions across the Americas.

Companies can give too much or too little access to data, applications and business processes, he told Unite AI. Agents may be given rights that are not congruent with what they are expected to do.

“The challenge is ensuring that an agent operates under the principle of least privilege, with access only to the data, systems, and actions required for a particular workflow,” Quintero said. That means granting access only to the data, systems and actions required for a specific workflow.

He identified clear data ownership, role-based access controls and continuous monitoring of interactions with internal and third-party systems as essential safeguards.

There is no such thing as safe local AI, Quintero emphasized.

Excessive permissions, weak encryption, insufficient monitoring and insecure integrations can make an on-device agent as risky as a cloud-based one – or even riskier.

Security is not just about where a system’s model is running; it’s also about how it’s configured and managed.

The New Boundary Is Permission

The question for the ordinary user is whether an AI agent’s access is appropriate for its task.

An agent should not have to scan an entire drive if it is only asked to summarize one document. An organizing tool for photos should not automatically read financial data or be able to send messages. Permissions must be tight, transparent and revocable.

The agent-account and workspace model in Microsoft is a step towards the separation of an agent’s activity from a person’s session. However, the protections need to be effective beyond the demonstration.

Microsoft’s Windows agent-security guidance calls for user authorization, logs of activity and a mechanism to approach that limits agents to the access they need. Employers and developers need to monitor the agents’ reaction to misguided instructions in documents, websites, or messages.

During an authorized task, agents can come across untrusted content. When an agent sees hostile text as a command, it could be a big security issue because the information can be easily accessed.

Human approval and least-privilege access must go hand-in-hand with containment for sensitive actions.

Convenience Needs Clear Limits

The AI PC will not only be evaluated on the amount of work agents can automate, but also on the user’s trustworthiness to share personal and professional data.

While local processing, the agentic AI can enhance speed and minimize certain data exposures, it doesn’t remove the need for permissions and oversight.

The principle should be straightforward: agents should only have access to what is necessary for the job, and users should be able to comprehend, track and block what an agent does.

Control must not be an afterthought, as computers become more capable of acting for us. It needs to be within the design.

Adlin Pertishya is a technology and business reporter who has worked for Finance Outlook India, Silicon India, StartupBeat and others.