Industry executives and experts share their predictions for 2026. Read them in this 18th annual VMblog.com series exclusive.
By Pranava Adduri, Bedrock Data CTO and co-founder
The transition from generative AI to agentic AI marks a fundamental shift in enterprise infrastructure. Unlike chatbots designed to generate text, autonomous agents execute workflows, interact with critical business systems and make decisions that impact operations in real time. Microsoft and NVIDIA’s 2026 roadmaps make this direction unmistakable. As security teams prepare for this transition, they face an important realization that the perimeter-focused security models of the past decade no longer suffice. Organizations must now secure the logic and data that drive business decisions themselves.
This transition introduces novel points of vulnerability that traditional security approaches can’t address. Here are three predictions that will reshape enterprise data strategies in 2026.
The Rise of the “Autonomous Insider”
Security teams must distinguish between a human user and an agent user. If you fail to do this, you invite high-speed insider threats where an agent acts on a hallucination or prompt injection. It will compromise systems faster than a human SOC can react.
Unlike traditional employees bound by cognitive limitations and institutional knowledge, agents operate at machine speed. An AI agent accessing an ERP system can execute hundreds of transactions in seconds. An agent with access to a codebase can modify repositories faster than a development team can review changes. An agent with financial credentials can move capital across accounts instantaneously. The “Autonomous Insider” threat vector is not hypothetical; it is the inevitable result of treating agents as tools rather than recognizing them as digital employees with material business impact.
Organizations that treat agents as simple interfaces to existing systems will find themselves exposed to compromise at a velocity that existing security controls cannot match. The solution requires a fundamental rearchitecture of identity and access controls to recognize agents as distinct actors within the security model.
RAG is a Data Supply Chain and Leak Point
The debate between fine-tuning and RAG is over. Retrieval-Augmented Generation won because agents require real-time context. However, RAG pipelines function as data supply chains. If your retrieval layer fetches a sensitive document because it was technically accessible to the user, the agent will process it. This problem only magnifies as we onboard multi-modal RAGs given foundational model advancements: data such ID verification images and security footage are ripe with sensitive data that are also notoriously hard to detect.
The retrieval layer is now the most critical control point in the stack. We must move from model security to context security. Teams must ensure data is classified and filtered before it enters the agent’s context window.
This shift demands a rethinking of data classification and access control. Historical approaches treated these as separate problems. Classification answered “what is this data?” while access control answered “who can see this?” In the agentic era, these must become a unified function: “what data should this agent see, given its purpose and the sensitivity of the information?”
Without this integration, sensitive documents slip into agent context windows not through malicious intent but through the simple architecture of systems designed for human-scale interactions.
Machine Volume Kills Manual Policy
We are entering an era of machine-generated data volume. When agents begin generating code, documentation, and datasets, the volume of enterprise data will increase exponentially.
Manual data classification already struggles to keep up with human output. It will fail completely against machine output. Security must rely on automated data lineage. If an agent reads source data marked confidential, its output must automatically inherit that classification. If you cannot automate the labeling, you cannot govern the file.
This prediction strikes at the operational foundation of data governance. Compliance and security teams have long relied on human judgment to make classification decisions. In the agentic era, this simply will not scale. Governance frameworks must become self-executing, with classification inheritance built into data pipelines themselves.
Preparing for the Agentic Transition
The agentic era demands a new security mindset. The move from generative to agentic AI is not a marginal evolution. It represents a change in how business systems operate and how security must evolve to protect them. Organizations that prepare now, by rearchitecting identity models, unifying classification with access control and automating governance, will emerge with resilient infrastructure. Those that delay will face a widening gap between the velocity of agent-driven business processes and the capability of security controls to govern them.
##
ABOUT THE AUTHOR
Pranava Adduri is CTO and co-founder of Bedrock Data, provider of the AI-native platform for data security, governance and responsible AI.
