In a report published October 15, 2026, NBC News reporter Joanna Stern said Microsoft CEO Satya Nadella wants Windows to become the safest and most powerful operating system for AI agents, while warning users not to trust those agents uncritically. The pairing puts a practical security question on the table: what can an agent access, and how can people oversee what it does? NBC News’s report
For related context, NeoTeo previously covered a proposed AI emergency brake attributed to Nadella.
What an AI-agent insider risk can mean
The term covers several different situations, not just deliberate wrongdoing. Identra distinguishes accidental exposure of sensitive data, intentional misuse of authorized access, a compromised account, and an agent manipulated by hostile instructions. To investigate an unusual action, Identra recommends establishing what data was involved, where it went, which instructions the agent received, the business purpose, and who owned the session. An unusual action is a lead to investigate, not proof of misconduct. Identra’s guide to AI insider risk
That distinction matters when a tool is approved for work. Approval of the tool does not automatically authorize every upload or instruction: the data, destination, permissions, and action still matter. Turning off model training also does not settle questions about whether an upload was authorized, how long content is retained, who can access the account, or whether content is shared.
Limit access and make actions auditable
Identra recommends giving agents and connected apps only the permissions needed for a particular task. Scope access to the folders, repositories, or mailboxes required, and treat reading and exporting as separate permissions. Require human approval before an agent shares information externally.
An audit trail should connect the user, account, agent, resource, action, and outcome. That makes it possible to examine an action in context rather than treating an agent’s unusual behavior as a verdict on the person or account behind it.
In a Bloomberg Tech interview published September 24, 2026, Ed Jennings, CEO of cybersecurity company Darktrace, said monitoring should look for agents behaving unnaturally or acting beyond their intended roles. Jennings also discussed Darktrace’s AI Signal Labs, placing his comments in the context of a security vendor’s work. His perspective adds a behavioral-monitoring angle to the access controls Identra recommends.
What BeyondID’s survey found
A BeyondID release dated June 25, 2025, reported results from the company’s 2025 survey of U.S.-based IT leaders. The figures describe organizations represented in that survey, not all organizations. BeyondID’s survey release
| Survey measure | Reported result | Population and period |
| Organizations saying they were ready for AI in security | 85% | BeyondID survey of U.S.-based IT leaders, 2025 |
| Organizations monitoring access or behavior for deployed AI systems | Fewer than 50% | BeyondID survey of U.S.-based IT leaders, 2025 |
| Organizations regularly mapping agents to critical assets | 30% | BeyondID survey of U.S.-based IT leaders, 2025 |
In this survey, self-reported readiness was more common than monitoring access or behavior, while fewer than a third of organizations regularly mapped agents to critical assets.
