Chapter 4
Operating and monitoring AI agents: Secure your data
Once identity is governed, data becomes the next — and often largest — risk surface for AI.
Data remains the primary target for attackers, and AI agents do not access data in fixed or predictable ways. They retrieve information dynamically from databases, documents, cloud storage, APIs, and external tools, often combining multiple sources to complete tasks. As agents analyze, summarize, redistribute, and otherwise act on this information, sensitive data may move through prompts, temporary memory, logs, outputs, and downstream workflows in ways that were never intended for or fully visible to security teams.
This fundamentally increases potential exposure and changes the data security equation.
Among the organizations participating in the research conducted by the Ponemon Institute for OpenText, 55% believe that AI agents increase the risk of data theft and 59% said AI makes privacy and data protection compliance more difficult.6 The challenge is no longer limited to protecting stored data — it is also governing how data is retrieved, processed, and shared across autonomous workflows.
6 Ponemon Institute, Managing Risks and Optimizing the Value of AI, GenAI & Agentic AI, March 2026.
of organizations believe that AI agents increase the risk of data theft.
Where data exposure actually occurs
In AI-driven environments, data exposure often happens through legitimate activity as opposed to overt breaches. Common exposure points include:
Retrieval pipelines that aggregate data from multiple repositories
Prompts and outputs that may unintentionally include sensitive data
Logs and telemetry that can capture confidential content for debugging or monitoring
Third-party integrations that may transmit data beyond enterprise-controlled environments
Training and fine-tuning workflows that may absorb sensitive data into model behavior
These pathways often fall outside the visibility of traditional security controls. Solutions such as data access governance help organizations identify sensitive repositories, understand who has access to them, and remediate excessive permissions before data is exposed.
A healthcare provider deploying an AI assistant to summarize patient records illustrates the risk. Even when access is authorized, protected health information could be exposed through prompts, logs, or downstream workflows in the absence of data masking or access controls, creating compliance and breach risk without any obvious failure event.
How traditional data controls fall short
Encryption, access controls, and perimeter defense remain essential but are no longer sufficient on their own in eliminating exposure.
In agentic AI environments, even encrypted data can still be accessed by unauthorized systems or actors if encryption keys are broadly accessible or if logs retain sensitive outputs. Sensitive information can still be decrypted by unauthorized systems or actors. Data protection must extend beyond encryption to include governance of key control, access policies, and usage monitoring.
The rise of retrieval-augmented generation (RAG) further complicates this landscape. AI agents increasingly retrieve contextual information from enterprise knowledge stores to improve accuracy and relevance. Without proper classification and filtering, these retrieval processes may surface confidential data beyond the agent’s intended scope.
What enterprises need to put in place
Securing AI at scale requires a data-centric approach that travels with the data itself.
Core practices include:
Classifying sensitive data and defining AI‑eligible versus restricted data sets
Applying masking, tokenization, or redaction before AI processing
Enforcing policy-based access controls for data retrieval and usage
Maintaining control over encryption keys and monitoring key usage
Limiting retention of prompts, outputs, and temporary logs
Monitoring data flows across AI pipelines, integrations, and downstream systems
These controls reduce exposure while enabling AI systems to operate effectively.
For example, an AI system reviewing legal contracts may require access to legal language but not personally identifiable information. Masking sensitive fields enables analysis to proceed without exposing confidential details. Similarly, financial analysts can evaluate transaction patterns by using tokenized data rather than raw account details.
Data minimization plays a critical role. AI agents do not need full data sets to generate insights. Providing only the minimum required information reduces both exposure and blast radius.
of organizations say AI makes privacy and data protection compliance more difficult.7
7 Ponemon Institute, Managing Risks and Optimizing the Value of AI, GenAI & Agentic AI, March 2026.
Securing data to enable trusted AI at scale
Clear data boundaries also support compliance and regulatory requirements. Industries such as healthcare, financial services, and government must ensure that sensitive information is accessed, processed, and retained in accordance with policy — a unique challenge for AI systems that are operating dynamically.
Organizations increasingly map data lineage to understand where information originates, how it is used, and where it flows across AI workflows. This visibility enables security teams to identify exposure points early and enforce consistent governance.
Protecting sensitive data before it enters AI pipelines remains one of the most effective ways to reduce risk. When sensitive data is classified, minimized, and governed at the source, organizations are better positioned to scale AI while preserving trust, compliance, and operational resilience.
Which data is AI-eligible vs. restricted
How sensitive fields are masked or tokenized
Who controls encryption keys used by AI systems
How long prompts, outputs, and logs are retained
How data movement across AI workflows is monitored
To operate AI agents safely at scale, organizations need a data-centric security approach that governs how sensitive information is accessed, processed, and shared across AI workflows. OpenText recommends starting with data discovery and classification to determine what data is appropriate for AI use and what must remain restricted.
Once identified, sensitive data should be protected with techniques such as masking, tokenization, or encryption before it enters AI pipelines. Organizations should also enforce policy-based access controls and strong encryption governance to ensure that only authorized systems and users can interact with sensitive information.
Because AI systems dynamically access data across repositories and applications, organizations must maintain continuous visibility into where sensitive data resides and how it is accessed. Monitoring access patterns, enforcing security policies, and protecting data at the source help reduce exposure, support regulatory compliance, and enable enterprises to scale AI adoption with confidence.
Operate AI securely with OpenText data security solutions.
