Cyberhaven is making a timely bet on where enterprise AI risk is actually heading. The company’s new Agentic AI Security offering is designed around a problem that has been building faster than many governance teams seem willing to admit: AI is no longer confined to a browser tab, a chatbot window, or a controlled SaaS workflow. It is beginning to act more like software with initiative, operating on endpoints, accessing data, using tools, and executing tasks with a degree of autonomy that changes the security equation. That matters because much of the first wave of AI governance was built around prompts, not actions. It focused on who used ChatGPT, what they pasted into Gemini, and whether sensitive information was being exposed through mainstream web interfaces. Useful, sure, but increasingly incomplete.
Cyberhaven’s framing centers on what it calls “shadow agents,” meaning AI systems running outside formal enterprise visibility and control. That idea lands because it reflects a broader shift already underway inside organizations: employees are not just experimenting with generative AI tools anymore, they are beginning to deploy agents, assistants, and endpoint-based automations that can interact with local files, internal tools, development environments, and connected services. Once that happens, the old monitoring model starts to look thin. It is one thing to inspect prompts sent to a cloud chatbot. It is another to understand what an autonomous agent running on a laptop is reading, what systems it touches, what commands it triggers, and what data it moves in the process.
The company backs that argument with figures from Cyberhaven Labs claiming that enterprise adoption of endpoint-based AI agents grew 276 percent over the past year, outpacing the growth of GenAI SaaS tools by more than three times. It also points to the rise of endpoint coding assistants, which reportedly jumped from 20 percent adoption to 50 percent in 2025. Even allowing for the usual caution around vendor-linked research, the directional point is hard to miss. AI usage is becoming more embedded, more operational, and more decentralized. The risk is not just that employees may share sensitive information with a public model. The risk is that agentic systems, often assembled quickly and adopted informally, are being granted meaningful access to enterprise environments before security teams have a clear inventory of what exists, let alone how it behaves.
That is where Cyberhaven wants to define a new control layer. Its Agentic AI Security launch expands the company’s broader AI and data security platform around three functions: visibility, observability, and control. In practical terms, the visibility piece is about discovering AI agents, MCP servers, and related endpoint connections. Observability is about seeing behavior rather than just presence, including what data agents access, what tools they invoke, and how execution flows unfold. The controls layer is the most important one, really, because it aims to enforce guardrails in real time while the agent is operating, not afterward when the data is already gone or the action has already been taken. That distinction is crucial. If AI is executing work rather than merely suggesting it, post-event logging is not enough.
The sharper message beneath the product announcement is that the endpoint is re-emerging as the decisive terrain of enterprise security in the AI era. For a while, the center of gravity seemed to be shifting toward SaaS governance, browser monitoring, and API-level oversight of foundation models. Cyberhaven is arguing that those approaches now see only part of the picture, and perhaps not the most consequential part. If organizations continue to rely on SaaS-only visibility while Agentic AI systems proliferate locally across employee machines, developer environments, and business endpoints, they may end up governing the visible fringe while the real execution layer remains largely opaque. That is not a great place to be.
Nishant Doshi, Cyberhaven’s CEO, put it plainly: AI is moving from generating content to executing work. That line gets at the real significance of the launch. The concern is no longer limited to what users ask AI to do. It is increasingly about what AI does next, with what permissions, against which data, and under whose authority. That shift moves security from a content-filtering mindset to an execution-governance mindset. Slightly awkward phrase, maybe, but the idea is real. Enterprises do not just need policies for AI use. They need operational controls for AI behavior.
Cyberhaven’s announcement also reflects a broader market scramble to define the next category in enterprise AI security. Everyone in the sector can see the same pattern: copilots are evolving into agents, local models and endpoint integrations are becoming more common, and the neat boundary between user action and machine action is eroding. Vendors are now racing to decide whether the future control point is the browser, the API, the identity layer, the cloud workload, or the endpoint. Cyberhaven is clearly betting that autonomous AI running close to the user, close to the data, and close to real enterprise workflows will become too important to monitor indirectly.
The launch, then, is not just a feature extension. It is a statement about where security teams should be looking next. Shadow AI was already a problem when it meant unsanctioned chatbot use. Shadow agents are more serious because they are not merely producing text or code suggestions. They are positioned to take action. Once that becomes normal across the enterprise, visibility alone is not enough, and retrospective analysis is not enough either. Security has to move to the point of execution. Cyberhaven sees that shift coming and wants to own the category built around it. Whether the market adopts its exact terminology is another question, but the underlying issue is real, and it is arriving fast.