AI Agents Move Onto Windows PCs, and Permissions Become the New Battleground
Microsoft security chief Hayete Gallot warned on October 19 that overprivileged AI agents threaten enterprise data as NVIDIA, Perplexity and Microsoft ship local agents for Windows PCs.

Why it matters
- Microsoft security EVP Hayete Gallot warned on October 19 that overprivileged 'agentic insiders' can access sensitive data and systems.
- NVIDIA RTX-powered Windows PCs now support locally running AI agents; Perplexity has expanded agent capabilities to compatible Windows computers.
- 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 limits.
Microsoft's security chief warned on October 19 that AI agents running on corporate PCs are frequently overprivileged, able to reach sensitive data, tools and systems while making decisions inside critical business processes. The warning from Hayete Gallot, Microsoft's Executive Vice President of Security, lands as agents stop being a forecast and start shipping: NVIDIA's RTX-powered Windows PCs now support locally running AI agents, Perplexity has expanded its agent capabilities to compatible Windows computers, and Microsoft is building safeguards to control how agents access resources and take action.
The shift matters because it moves AI from the chat window into the operating system, the applications and the files people use every day. New Windows PCs and software released in recent months introduce multistep autonomous workflows for ordinary users and for enterprise environments. Each of those workflows carries permissions, and each permission carries risk.
What did Microsoft's security executive actually say?
Gallot told organizations they must do more to secure what she called "agentic insiders" — agents that hold credentials, read documents and act inside business systems with less scrutiny than a human employee would face. Her framing, reported by Unite AI, reflects a concern that an agent's error can have repercussions well beyond a wrong answer. A mistake can reveal sensitive information, alter critical files or trigger unwanted activity inside business applications.
Microsoft's Windows security strategy rests on containment, limited permissions and limited oversight. The company is also developing mechanisms to govern how agents approach resources and take action on a user's behalf. The stakes are straightforward: an agent that can summarize a document can, if overprivileged, also read the drive around it.
Does local AI processing guarantee privacy?
No, and this is the central claim of the security argument now forming around the AI PC. Running a model on-device reduces the amount of sensitive data transferred to cloud services, but local processing is not a safety net. An on-device model can still access too many files, run with poor security configuration or hold unsafe links to other applications.
"There is no such thing as safe local AI," said Bayron Quintero, CEO and co-founder of iData Global, a company specializing in AI solutions across the Americas. Excessive permissions, weak encryption, insufficient monitoring and insecure integrations can make an on-device agent as risky as a cloud-based one — or riskier, he argued in comments to Unite AI.
Quintero located the problem in organizational governance rather than in the model itself. Companies can give too much or too little access to data, applications and business processes. Agents may receive rights that do not match 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.
He identified three safeguards as essential:
- clear data ownership, so someone is accountable for what an agent touches;
- role-based access controls, so permissions follow the task rather than the agent;
- continuous monitoring of interactions with internal and third-party systems.
His conclusion reframes the debate: security is not about where a model runs. It is about how the system around it is configured and managed.
Why is permission the new security boundary?
For the ordinary user, the operative question is whether an agent's access fits its task. An agent asked to summarize one document should not scan an entire drive. A photo-organizing tool should not read financial records or gain the ability to send messages. Permissions, the argument goes, must be tight, transparent and revocable.
Microsoft's agent-account and workspace model represents a step toward separating an agent's activity from a person's session. The open question is whether those protections hold beyond demonstrations and lab conditions.
Microsoft's Windows agent-security guidance calls for three things:
- user authorization before agents take sensitive actions;
- logs of agent activity that users and administrators can inspect;
- an approach that limits each agent to only the access a task requires.
The guidance also addresses a subtler threat. During an authorized task, an agent can encounter untrusted content in documents, websites or messages. If an agent reads hostile text as a command — a prompt injection — the consequences can be serious, because the agent already holds access to the information and systems around it. Employers and developers, the guidance says, must monitor how agents react to misguided instructions embedded in the material they process.
The remedy is structural, not incidental: human approval and least-privilege access must operate together with containment for sensitive actions.
What is at stake for the AI PC market?
The AI PC will be judged on more than how much work its agents can automate. It will also be judged on whether users trust it with personal and professional data. That trust, according to the argument running through both Microsoft's warnings and Quintero's analysis, depends on visible control rather than on processing location.
Local, agentic AI improves speed and reduces certain categories of data exposure. It does not remove the need for permissions and oversight. A user who cannot comprehend, track and block what an agent does has no meaningful control over it, regardless of whether the model runs in a data center or on the machine in front of them.
The principle is simple to state and hard to implement: agents should access only what the job requires, and users should be able to see, audit and revoke that access. The comparison that defines the moment is the one Quintero drew — an overprivileged on-device agent can be riskier than a cloud service, because it sits inside the perimeter it threatens.
Control must not arrive as an afterthought, the argument concludes, as computers become more capable of acting on their users' behalf. It has to sit inside the design from the start — in the permission model, in the logs, and in the approval gates that stand between an autonomous agent and the files it was never meant to read.
Source: unite.ai
More from Rebecca Stone
Show full bio
Correspondent covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.
728 articles
Related articles
- Tanium launches AgentSight to track and govern AI tools on endpoints
- AI Security in 2027: Agent Permissions Will Define the Blast Radius
- BeyondTrust Predicts First Autonomous AI Breach by 2027
- OpenAI adds Lockdown Mode and risk labels to ChatGPT
- Microsoft calls AI agents insider threats and ships Guardian Agents