In Brief (TL;DR)
AI adoption is accelerating, but governance and security controls are struggling to keep pace. Traditional Data Loss Prevention (DLP) focuses on preventing sensitive information from leaving the organisation, but AI introduces new risks, including data inference, synthesis, memory retention, and autonomous actions by AI agents. Effective DLP for AI must govern the entire data-to-action chain, from user inputs and data retrieval through to AI outputs and agent actions. Organisations that successfully scale AI will be those that combine strong identity controls, data governance, AI security, and business oversight, enabling innovation without creating new security, compliance, or operational risks.
The Model Is Not The Data Boundary
Most AI security conversations start with a simple question:
How do we stop employees from pasting sensitive data into public AI tools?
It's an important question, but it's no longer the most important one.
Modern AI systems do far more than process prompts. They retrieve information from enterprise repositories, combine data from multiple sources, retain memory, generate new insights, and increasingly take actions on behalf of users.
As a result, DLP for AI must evolve beyond monitoring data movement. It needs to govern the entire journey from data access to AI-driven action.
introducing AI capabilities faster than governance processes can keep up.
The risk isn't simply what data enters an AI model. It's what the AI can do with that data afterwards.
Why So Many AI Projects Fail
Research from RAND has highlighted that many AI projects fail, often at a higher rate than traditional IT initiatives.
The problem is rarely the AI model itself.
Organizations often struggle because they lack:
- Clear business objectives
- High-quality, accessible data
- Appropriate governance
- Scalable infrastructure
- Realistic expectations
In other words, AI failures are frequently enterprise readiness failures, not technology failures.
AI Adoption Is Moving Faster Than Governance
Organizations are embracing AI at unprecedented speed.
Business teams are experimenting with AI assistants. SaaS vendors are embedding AI into products. Developers are building AI-powered workflows and agentic systems.
Yet governance is struggling to keep pace.
This gap creates the conditions for:
- Shadow AI
- Data leakage
- Over-permissioned AI agents
- Compliance concerns
- Unscalable deployments
The challenge isn't adoption. It's control.
AI isn't waiting for governance to catch up. Governance must learn to operate at AI speed.
AI Changes the Meaning of Data Loss
Traditional data loss prevention focuses on identifiable events, such as uploading a confidential document or emailing sensitive information outside the organization.
AI introduces entirely new risks.
Data Synthesis
AI can combine multiple low-risk data sources to create highly sensitive conclusions, exposing insights that were never stored in a single location.
Data Inference
Models can derive confidential information that was not explicitly contained within the original data.
Retrieval Amplification
An AI agent can access and analyze information at a speed and scale impossible for a human user, dramatically increasing potential exposure.
Data Transformation
Sensitive data can be transformed into summaries, code, recommendations, reports, or visualizations that evade traditional controls.
Persistence
Information may remain within conversation histories, model memory, vector databases, embeddings, or observability platforms long after the original interaction.
Action
The greatest shift is that AI can move beyond providing information and begin acting on it by sending messages, updating records, deploying code, or initiating transactions.
This means data loss is no longer just about information leaving the organization.
It can also occur when AI uses information in ways that were never intended.
The EchoLeak Lesson
The EchoLeak vulnerability (CVE-2025-32711) highlighted a critical reality of enterprise AI security.
AI systems sit between trusted internal data and untrusted external content. They consume emails, documents, websites, chat messages, code repositories, and third-party information.
If a model cannot reliably distinguish between data and instructions, attackers may be able to manipulate AI behaviour through content the system is authorised to access.
The broader lesson is clear:
DLP for AI cannot rely solely on content inspection.
Organizations must also understand:
- Prompt injection risks
- Data lineage
- Identity context
- Retrieval permissions
- Tool usage
- Output destinations
Securing the Entire AI Data Lifecycle
Effective AI governance must protect every stage of the AI lifecycle.
Rather than applying controls only at the point of data entry, organizations should focus on:
- User inputs and data sensitivity
- Model selection and hosting location
- Retrieval permissions
- Prompt construction
- Vector databases and embeddings
- AI memory and retention policies
Training and fine-tuning datasets - AI outputs
- Tool usage and automation
- Logging and monitoring
This broader approach forms the foundation of modern AI data security.
Why Identity Context Matters
Not every user, agent, model, or business process should be able to access or use information in the same way.
A mature AI security strategy considers:
- Who initiated the request
- Which AI agent is acting
- Which model is being used
- Why the data is being accessed
- What action is being performed
- Where data will be stored or shared
The same dataset may be appropriate for one business process but completely inappropriate for another.
This is where identity security and data security must converge.
Data classification tells you what the data is. Identity context tells you whether its use is appropriate.
Why Traditional DLP Is No Longer Enough
Traditional DLP was built to identify patterns, keywords and file movement. AI changes this model entirely.
Sensitive information can now be summarised, inferred, transformed into code, fragmented across multiple interactions, or used by an AI agent to trigger actions.
This means organisations need controls that understand not just data, but also context, identity, purpose and intended outcomes.
Traditional DLP protects data in motion. DLP for AI must govern data in use.
Building an AI Data Control Plane
To secure enterprise AI, organisations need an integrated control plane that combines:
- AI discovery and approved model governance
- Data classification and semantic DLP
Identity governance and privileged access controls - Prompt and response inspection
- Vector database and retrieval security
- Data residency, retention and privacy controls
- Agent authorisation and runtime monitoring
Together, these capabilities govern the entire data-to-action chain, helping organisations scale AI safely without slowing innovation.
Governed Adoption, Not AI Avoidance
Many organizations respond to AI risk by blocking access altogether.
That rarely works.
Users simply move to personal accounts, browser extensions, or unsanctioned tools, creating more risk and less visibility.
The goal should be governed adoption, where secure AI experiences are accessible, productive and aligned with business needs.
Final ThoughtÂ
Ready to make AI Accountable?
Build governance into every agent identity.
Frequently asked questions
What is DLP for AI?
DLP for AI extends traditional data loss prevention by monitoring not only data movement but also how AI systems retrieve, infer, transform, retain, and act on information.
Why isn't traditional DLP enough for AI?
Traditional DLP focuses on patterns and file transfers. AI can summarize, infer, and transform data in ways that bypass conventional controls.
What is Shadow AI?
Shadow AI refers to employees or teams using unapproved AI tools, models, or services outside established governance and security controls.
Why is identity important in AI security?
Identity provides the context needed to determine whether an individual or AI agent should access specific information and what actions they are authorized to perform.
How can organizations secure AI adoption?
Organizations should implement a comprehensive AI governance framework incorporating identity security, data classification, DLP for AI, model controls, retrieval governance, monitoring, and runtime enforcement.
