Innovation & Tech

On-Device AI Agents Are Here — and Local Doesn't Mean Private

Microsoft's security chief warns of 'agentic insiders' as NVIDIA, Perplexity and Microsoft push local AI agents onto Windows PCs. Local processing isn't privacy.

By Nathan Brooks

3 min read

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

What's News

  • On October 19, Microsoft security EVP Hayete Gallot warned organizations to secure overprivileged 'agentic insiders'.
  • NVIDIA's RTX-powered Windows PCs now support locally running AI agents; Perplexity has expanded agents to compatible Windows machines.
  • iData Global CEO Bayron Quintero: 'There is no such thing as safe local AI.'
  • Microsoft's Windows agent-security guidance requires user authorization, activity logs and least-privilege access.

AI agents have moved out of the chat window and into the operating system, and on October 19 Microsoft's Executive Vice President of Security, Hayete Gallot, warned that organizations must do more to secure what she calls "agentic insiders." Her point: agents are frequently overprivileged, with access to sensitive data, tools and systems while making decisions in critical processes.

The shift is no longer a forecast. NVIDIA's RTX-powered Windows PCs now support locally running AI agents. Perplexity has expanded its agent capabilities to compatible Windows computers, enabling workflows that touch local files and connected applications. Microsoft, meanwhile, is building safeguards to control how agents access resources and take action.

That's the practical question for anyone deploying this stack: what does local processing actually change? Less than you'd hope.

Does local AI mean private AI?

Not automatically. Running an agent on your PC can cut the amount of sensitive data flowing to cloud services, but local processing is not a security blanket. An on-device model can still access too many files, run with poor security or hold unsafe links to other applications.

Bayron Quintero, CEO and co-founder of iData Global, a company specializing in AI solutions across the Americas, puts the blame on configuration rather than the model itself. Companies can give too much or too little access to data, applications and business processes, he told Unite AI. Agents may receive rights that don't match their actual job.

"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.

His prescription for practitioners breaks down into three safeguards:

  • Clear data ownership.
  • Role-based access controls.
  • Continuous monitoring of interactions with internal and third-party systems.

Quintero's bottom line is blunt: there is no such thing as safe local AI. Excessive permissions, weak encryption, insufficient monitoring and insecure integrations can make an on-device agent as risky as a cloud-based one — or riskier. Security depends on how a system is configured and managed, not on where the model runs.

What does this change for IT teams?

The new security boundary is permission, not location. The test for any deployment is whether an agent's access fits its task. An agent asked to summarize one document shouldn't scan an entire drive. A photo-organizing tool shouldn't read financial data or be able to send messages. Permissions must be tight, transparent and revocable.

Microsoft's agent-account and workspace model is a step toward separating an agent's activity from a person's session. But those protections need to work beyond the demo stage.

Microsoft's Windows agent-security guidance calls for three things: user authorization, logs of activity, and an approach that limits agents to the access they need. Employers and developers also need to watch how agents react to misguided instructions embedded in documents, websites or messages.

Here's the sharpest new risk. During an authorized task, an agent can encounter untrusted content. If the agent reads hostile text as a command, that becomes a serious security issue — because the agent already holds the keys to accessible information. Microsoft's broader Windows security strategy rests on containment, limited permissions and limited oversight, driven by concern that an agent's error could ripple beyond a wrong answer. It could reveal sensitive information, alter critical files or trigger unwanted actions inside business applications.

Convenience needs clear limits

The AI PC will be judged not just on how much work agents can automate, but on whether users trust them with personal and professional data. Local, agentic AI buys speed and reduces certain data exposures. It does not remove the need for permissions and oversight.

Human approval and least-privilege access must go hand in hand with containment for sensitive actions. The principle is simple: agents get access only to what the job requires, and users can comprehend, track and block what an agent does at any moment.

As computers become more capable of acting for us, control can't be an afterthought. It has to sit inside the design from day one.

Source: unite.ai

Share this article:

More from Nathan Brooks

Nathan Brooks

Show full bio

News editor covering regulation, policy and their impact on entrepreneurship, startups and venture funding at Business Bearings.

2,232 articles

Related articles

« Previous articleNext article »