AI-ready by design: all the context AI needs, none of the risk
Centrally control what thousands of agents and queries can access, with a new policy engine and protection enforced at the data, agent and inference layers.
Every AI project runs on context: the data, documents and company knowledge that people and agents draw on to make decisions and take action. And the context that makes AI most useful is usually the most sensitive, from customer records and health information to financial data and contracts.
That leaves most enterprises with a trade-off. They can give AI enough context to be useful and accept the exposure, or hold sensitive data back and delay or diminish their AI initiatives. EMA’s State of AI Friction research found that 83% of organizations have had AI projects delayed by security or compliance reviews, 82% have shipped AI with capabilities deliberately limited, and 69% are running AI agents under a protection model that wasn’t built for them. Together, that’s the AI friction tax, and you’re already paying it. The friction isn’t in the AI, though. It’s in the data that powers it.
Protegrity protects the data, agents and inference behind enterprise AI. For more than two decades, we’ve protected sensitive data for some of the world’s largest brands, and today our solutions protect the privacy of more than a billion people. Our secure context architecture ends the trade-off: it gives every agent exactly what it needs, and controls what it can see and do.
With our latest release of Protegrity AI Team Edition, available today, we’re beginning a multi-phase plan to secure every aspect of context in the AI era. At its core is a new policy engine that gives security teams a single solution for applying data and AI security policies to every person, agent and query: one engine, enforced at every layer. Our roadmap will continue to expand that control across the context layer, agent intent and how agents communicate with each other.
Secure context at every layer
AI security comes down to three places where protection must hold, and each one fails in its own way.
In the data: protection applied to the values themselves travels with them. Protected where they live, sensitive values stay protected through every pipeline, lakehouse and model they pass through, without rebuilding controls system by system.
At the agent: agents work by passing context, from user to agent, agent to model, agent to agent and agent to data. Each step should get only the context its task needs and act only within policy, and that has to be decided at runtime, not when access was granted.
At inference: not all sensitive data starts in a protected table. People paste it into prompts, and models can repeat it in answers. Every prompt and response needs its own checks, for sensitive data and for what’s being asked and answered, applied the same way across every model and app.
All three layers run on one shared core, which is why they hold together. The Protegrity PII model classifies sensitive data wherever it appears. The policy engine decides what each person, agent and query can see and do, and in what form, and applies those policies across every layer. The same protection methods, from tokenization and encryption to masking and redaction, carry out those decisions whether the data sits in a table, a query result, a prompt or an answer. Data & AI Fast Lanes gives security teams control over all of it in a single solution. And because the core stays the same from one AI project to the next, security can review it once and reuse that review instead of starting over with every new use case.
What’s new in this release
New policy engine
The new policy engine is built on Cedar, the open-source policy language, and it’s extensible to agents and inference, not just data. It’s tightly integrated with the Protegrity PII model, so visibility turns straight into enforcement. The same policies hold across data, agents and inference instead of being rebuilt tool by tool, so new AI projects start from controls that already exist.
Protegrity PII model
Every protection decision starts with knowing what’s sensitive. Our fine-tuned PII detection model reaches 0.907 precision and 0.915 recall — higher scores than OpenAI’s Privacy Filter, Amazon Comprehend, Microsoft Presidio and GLiNER. It recognizes 69 entity types across languages, including health conditions, biometric data and crypto wallet addresses, which most detectors don’t support. To test it, we built a new benchmark dataset with about 13 labeled entities per document, 1.6 times the density of the next-richest dataset and the fewest unlabeled PII values of any we examined. Fewer misses means less sensitive data reaches a model. Fewer false flags means more data stays usable instead of going back behind a review.
Data & AI Fast Lanes
Data & AI Fast Lanes is a new UI and user experience for security teams, starting with detection, classification, and protection policies. Security teams can see where sensitive data lives, decide what happens to it, and sanitize or redact sensitive values in text before any model sees it. It takes the friction out of saying yes to AI, so more prompts, documents, tickets and transcripts are ready for AI sooner.
AI Gateway
AI Gateway provides central, extensible guardrail orchestration for every prompt and response. Run Protegrity’s sensitive data, semantic and behavioral guardrails alongside the ones you already use, and plug in new ones as you need them. You keep the guardrails you’ve invested in without wiring each AI application separately, every prompt and response passes the same checks whichever app or agent sent it, and security reviews one control point instead of one per app.
Secure Query (Tech Preview)
Secure Query runs transparent, high-performance search, sort, filter and aggregates such as min, max and sum directly on protected data. It adds attribute-based access control (ABAC), row-level filtering and runtime protection that doesn’t require protecting the source data first — with no database triggers or user-defined functions to deploy. Sensitive datasets that sat behind a review can feed analysts and agents without building de-identified copies or re-engineering the source, and each request returns only what that user or agent is allowed to see.
Iceberg Protector (Tech Preview)
Apache Iceberg lets many compute engines and applications work on the same tables. Iceberg Protector gives that ecosystem a secure foundation, so Protegrity protection holds whichever engine queries the table. That makes your data fabric ready for enterprise AI and analytics workloads, without rebuilding controls engine by engine or keeping sensitive tables out of the lakehouse.
What this means for you
AI ships sooner, because security reviews don’t start over with each project. AI does more, because it works with your most sensitive data instead of a partial or watered-down dataset. And agents stay within policy, because their activity is checked as it happens.
That’s AI-ready by design.
Availability
Available now in the latest release of Protegrity AI Team Edition: the new policy engine, the Protegrity PII model, Data & AI Fast Lanes and AI Gateway.
All of this runs on the Protegrity Provisioned Cluster (PPC), the Kubernetes-based foundation that AI Team Edition deploys onto. PPC runs on Azure and AWS, so you can run AI Team Edition in the cloud you already use, on managed Kubernetes your infrastructure teams already know. It scales as workloads grow without re-architecting, and individual components can be updated independently. Because these capabilities share one cluster, you deploy and operate one environment instead of stitching separate products together, and new capabilities land on infrastructure you already run. Our roadmap includes extending support to Google Cloud and to on-premises environments such as Red Hat OpenShift.
In tech preview: Secure Query and Iceberg Protector. Tech preview doesn’t mean prototype. Both are fully functional releases that customers are using today. We’re bringing new capabilities to market this way so those customers can shape the final version before it’s generally available.
See it for yourself
Request a demo to see the new release in action, or explore Protegrity’s product overview for more details.
For the full findings behind the AI friction tax, read EMA’s State of AI Friction report.