What happened
Cyberhaven Chief Product Officer Manish Kalia announced Agentic Data Security on November 5, 2026. The company describes ADS as a new product in Cyberhaven Flow, its platform for data security and insider-risk protection. The announcement frames the product around AI apps and agents running across enterprise endpoint environments, including desktop applications, coding tools, and browser sessions. Cyberhaven says ADS was available for purchase on the day of the announcement. The supplied source does not give a price, deployment prerequisites, or a customer eligibility breakdown.
The product is designed to help security teams discover AI tools and agents used across endpoints, including unsanctioned applications. Cyberhaven says ADS inventories agentic applications and can identify models, plans—free or enterprise—extensions, and tools they use. It also says the system monitors MCP tool calls, whether they operate over stdio or a network. The purpose, as presented by the company, is to make agent use visible to security teams before they decide how to govern it.
ADS also records an agent session as a workflow. According to the announcement, investigators can review prompts, responses, tool calls, file reads and writes, commands, and data destinations, with data lineage showing the data involved and how it was transformed. Cyberhaven says its system attributes actions to the person or system behind them. This is intended to let security teams follow the movement of information through interactions among people, agents, tools, and destinations rather than inspecting only one channel or application at a time.
The company says ADS is built on Cyberhaven Flow’s existing data-lineage, data-loss-prevention (DLP), data-security-posture-management (DSPM), and insider-risk capabilities. Its stated objective is to extend existing protection to agentic workflows and let teams act when a workflow poses risk. The supplied material describes warnings or blocking of sensitive data before it leaves, alongside controls at runtime on endpoints. It does not specify the complete policy configuration process or provide technical performance figures.
Cyberhaven argues that separate controls for human and agent activity can miss risks at handoffs, because employees and agents may work on the same information in one workflow. It says endpoint-based adoption of agentic AI apps doubled between 2025 and 2026, citing Cyberhaven Labs. The announcement presents that increase as part of the motivation for the product, alongside customer concerns about agents accessing and exfiltrating sensitive data or performing harmful actions, intentionally or by mistake.
What the archive adds
This is a new Cyberhaven product announcement in a developing enterprise-agent security thread. The archive previously covered Adversa AI’s platform for securing coding agents, which monitors agent calls, tool executions, and endpoint actions. ADS is a distinct launch: Cyberhaven positions its module around data lineage and extending its existing data-security controls to people-and-agent workflows. The earlier coverage makes this a new fact on a known story, not a restatement.
Trend signal
The launch supports agent security moving to runtime boundaries. Cyberhaven says ADS observes agent activity on endpoints and can warn or block sensitive-data movement at runtime, extending controls beyond instructions given to a model. This is a vendor’s description of its product, not independent evidence that such controls reliably stop exfiltration in production.
Why now
The announcement links the launch to growing enterprise use of AI apps and agents. Cyberhaven Labs reports that endpoint-based adoption of agentic AI apps doubled between 2025 and 2026, and Kalia says security leaders are concerned about sensitive-data exposure and harmful agent actions. Those claims explain the company’s timing, but the supplied material does not define the adoption measure or provide its underlying data.
Connections to other events
The launch meets a broader enterprise-security question: how to govern agents that act with access inherited from the people or systems they serve. A recent Oracle AI Security Mesh announcement also brought AI capabilities into an enterprise security product, but its focus was vulnerability finding rather than agent data workflows. Cyberhaven’s stated distinction is that ADS follows data and actions across human and agent activity on endpoints.
Why it matters
For teams building coding agents, browser agents, or MCP-connected applications, enterprise controls may increasingly inspect the whole workflow: what data an agent reads, which tools it calls, what transformations it makes, and where the result goes. That has a practical design implication. An integration that works in a developer’s environment may encounter warnings or blocks when deployed under an organization’s data policies, even if the model interaction itself appears harmless.
ADS also illustrates a security posture in which the agent is treated as another actor in a data workflow, rather than as a special exception to policies designed for employees. If that approach becomes common, builders will need to make data destinations, tool permissions, and agent actions legible to the organization operating the endpoint. The announcement does not establish that Cyberhaven’s approach is becoming an industry standard, but it makes that governance layer more concrete for enterprise buyers and agent teams.
Limits and uncertainty
The source is a Cyberhaven product announcement and describes the company’s own capabilities and rationale. It does not provide independent testing, quantified detection or prevention results, customer deployment evidence, false-positive rates, or details about how ADS handles unsupported apps and tools. The adoption-doubling figure comes from Cyberhaven Labs, but the supplied material does not specify its sample, definition, or methodology. The source says ADS was available for purchase on November 5, 2026, but gives no pricing or packaging details. Claims about monitoring and blocking should therefore be treated as vendor statements rather than verified outcomes.
Skill to build now
Map one agent workflow that handles sensitive data: identify the files it can read or write, the tools and MCP servers it calls, and every destination for its outputs. This week, use that map to check whether your team can explain which identity initiated each action and where data leaves the workflow. The exercise is useful whether or not you adopt ADS: it makes the boundaries a data-security product would be expected to observe explicit.
Reader takeaway
Cyberhaven’s launch puts data lineage and endpoint controls at the center of its answer to agent risk: track what an agent does with sensitive information, then apply existing security policies to that activity. Builders should expect enterprise deployment reviews to look beyond model prompts and ask how agent tools, files, and destinations are governed.