Protegrity has updated its AI Team Edition platform with a unified policy engine that controls how people, AI agents and applications access data and interact with AI models.
Built on the open-source Cedar policy language, the engine provides a central point for applying security policies to data, agents, prompts and responses. It is integrated with Protegrity’s personally identifiable information detection model, allowing organizations to enforce access rules based on the sensitive data identified within an AI workflow.
“AI becomes more valuable when it has access to the context of the business, but that context is also where an organization’s most sensitive information lives,” said Michael Howard, CEO, Protegrity. “Enterprises shouldn’t have to choose between AI that’s useful and AI that’s secure. We’re building the secure context layer so they don’t have to: every agent gets exactly what it needs, and does nothing it shouldn’t. That’s what moves AI out of experimentation and into production.”
The release addresses the risks created when enterprise AI systems use internal documents, databases and company knowledge to answer questions or complete tasks. Those systems may also include agents that can write to databases, call external application programming interfaces and take actions without direct human involvement.
Research from EMA’s State of AI Friction report found that security or compliance reviews have delayed AI deployments at 83% of organizations. The research also found that 82% have released AI systems with reduced functionality because they could not resolve the associated risks. Another 69% are operating agents with the ability to write to databases and activate external APIs.
Protegrity’s architecture applies controls across the data supplied to AI systems, the permissions assigned to users and agents, and the prompts and responses passing through AI applications. The policy engine determines what each person, agent and query can access and which actions they are permitted to perform.
The updated platform also includes Protegrity’s fine-tuned PII detection model, which recognizes 69 types of sensitive information. Supported categories include health conditions, biometric data and cryptocurrency wallet addresses.
According to Protegrity, the model recorded precision of 0.902 and recall of 0.910. The company said those scores exceeded results from OpenAI’s Privacy Filter, Amazon Comprehend, Microsoft Presidio and GLiNER.
A new Data & AI Fast Lanes interface allows security teams to identify and classify sensitive information, establish protection policies and apply controls before data reaches an AI model.
Protegrity has also added AI Gateway, a centralized orchestration point that evaluates prompts and responses throughout AI workflows. Organizations can use the gateway to run Protegrity’s guardrails alongside existing security controls.
Two additional capabilities are available in technology preview. Secure Query allows applications to search, sort, filter and perform aggregate operations on protected data. It supports attribute-based access control and row-level filtering without requiring database triggers or user-defined functions.
Iceberg Protector extends data protection to Apache Iceberg environments, including the compute engines and applications that access data stored in the lakehouse. Protegrity said Secure Query and Iceberg Protector are fully functional and already being used by customers, whose feedback will inform the final versions before general availability.
The company plans to expand its security controls to cover additional elements of enterprise AI, including the context provided to models, agent intent and communications between agents.
The policy engine, PII model, Data & AI Fast Lanes and AI Gateway are available in the latest release of Protegrity AI Team Edition. The capabilities run on Protegrity Provisioned Cluster, a Kubernetes-based foundation that can be deployed on Microsoft Azure and Amazon Web Services.
Because the components operate on the same cluster, customers can deploy and manage them within one cloud environment. Protegrity plans to add support for Google Cloud and on-premises platforms such as Red Hat OpenShift.
