Summary: Microsoft’s first Cyber Pulse report, published February 10, 2026, says more than 80% of Fortune 500 companies had active AI agents built with Copilot Studio or Agent Builder. It also reports 29% of employees admitting use of unsanctioned agents and recommends zero-trust controls, agent registries, least privilege, telemetry, interoperability, and security. The article separates governance from security and calls both enterprise-wide responsibilities.
The first edition of Cyber Pulse arrived from Microsoft on February 10, 2026, under the byline of Vasu Jakkal, Corporate Vice President of Microsoft Security, and it stays with one question the whole way through: as AI agents settle into the enterprise in large numbers, how do security teams go about observing, governing, and protecting them? The opening figure comes from Microsoft’s own telemetry — more than 80% of Fortune 500 companies now have active AI agents in operation, built with low-code/no-code tools. The sample window covers the final 28 days of November 2025, and any agent created through Copilot Studio or Agent Builder falls inside the count.
The report earns its weight from two things: the scale of the data behind it and the timing of its release. Agents are past the demo stage; they carry real permissions, reach out to external APIs, and generate content on their own — production workloads in every sense of the word. Microsoft spoke at the moment enterprise agent counts had just crossed a threshold, and observability took first place on the security agenda as a result.
Where Agents Took Root: Heavily Regulated Industries Rank Near the Top
Cut the data by industry and the ranking looks like this: software and tech leads the pack at 16%, manufacturing takes second at 13%, financial institutions account for 11%, and retail stands at 9%. Manufacturing and finance — sectors with some of the heaviest regulation anywhere — both placing in the upper tier sends its own message: agent adoption left the startup world behind long ago and has pushed through the most conservative IT environments in existence.
The survey side draws on the 2026 Data Security Index: a questionnaire administered to 1,725 data security decision makers (fielded from July 16 through August 11, 2025), accompanied by ten in-depth interviews. Telemetry and questionnaire arrive at the same conclusion — adoption keeps climbing a steep slope while protection stays nearly flat.
A New Kind of Shadow AI: 29% of Employees Have Used Unsanctioned Agents
One more figure lands close to everyday work: 29% of employees admit they have used AI agents nobody approved to get work tasks done. Jakkal’s line will likely be quoted all year: “You can’t protect what you can’t see, and you can’t manage what you don’t understand.”
The risk an agent carries looks nothing like the shadow IT of earlier days. It inherits the permissions its creator held, acts on its instructions without oversight, and leaves its output where IT has no view — give one read access to a mailbox and you are looking at an attack surface that sends its own email and places its own orders. Security Brief was among the outlets passing along the report’s warning about ungoverned “AI double agents”: a security exposure and a compliance liability in the same package.
The Remedy the Report Prescribes: Zero Trust for Agents
The report’s central recommendation is to hold AI agents to the discipline already applied to employees and service accounts: grant the fewest permissions, verify explicitly, and assume compromise is a matter of time. Worked into practice, that becomes five foundational capabilities for governing agents:
- Registry: a single source of truth — start by inventorying every agent
- Access control: least privilege, without exception
- Visualization: real-time telemetry that keeps each agent action in view
- Interoperability: governance that runs across platforms rather than inside one vendor’s ecosystem
- Security: threat protection present from the start
Of the five, interoperability is the one that merits a second look. Microsoft does not bound governance within its own product line; it accepts outright that enterprise agents will arrive from many frameworks and vendors, and that governance has to span them all.
Governance Is Not Security — and Both Are Company-Wide Work
The report also takes pains to separate two words usually treated as interchangeable. Governance answers “who owns this agent and what may it do”; security covers enforcement and threat detection. Each completes the other, and neither stands in for the other — policy without detection stays on paper, and detection without policy wanders without direction.
A third message lands on organizational design: AI risk is an enterprise-level matter that cuts across functions, so legal, HR, compliance, and the board all belong at the same table instead of leaving the CISO to carry the load alone. Jakkal pitches security as momentum rather than fear: “Security is a catalyst for innovation.”
What Developers and Security Teams Can Do Right Now
Three moves are ready to make today. First, wire registry and telemetry interfaces into agents from day one — retrofitting visibility later only gets more expensive. Second, build the permission model on the service-account pattern: token lifecycles, narrowed scopes, and revocation at any time. Third, add an agent inventory exercise to the annual security plan — the same shift we followed in our 2026 opening outlook: the ground enterprise AI is fighting over is moving from “which model do we pick” to governance and observability. Microsoft has turned that piece of conventional wisdom into a number backed by its own telemetry; what is left is each company’s choice of when to act.