Bedrock Data Integrates With NVIDIA OpenShell to Bring Data-Aware Policy Enforcement to AI Agents

Bedrock Data brings enterprise data context to NVIDIA OpenShell, evaluating each agent action against data policy and returning the decision OpenShell enforces at runtime

SAN MATEO, Calif., September 28, 2026--Bedrock Data, the platform provider for data and AI security and governance, and data security posture management (DSPM), today announced it has extended Agent DLP™, its runtime data loss prevention for AI agents, to NVIDIA OpenShell, the open source agent runtime that is part of the NVIDIA Open Safety Platform. NVIDIA OpenShell gives autonomous agents a secure environment to run in with strong isolation and a zero-trust architecture where nothing is permitted by default. This integration offers an additional policy evaluation layer where Bedrock Data evaluates a request payload once the request is authorized by OpenShell. Bedrock Data's authorization layer evaluates the payload in the request against an enterprise's data policies. The integration is available today to Bedrock Data customers.

"The identities enterprises bind to agents are usually more permissive than the task requires, so an agent pursuing a goal can access data nobody meant it to have, combine sources into toxic data combinations and share sensitive data to finish the job," said Pranava Adduri, CTO and co-founder of Bedrock Data. "NVIDIA OpenShell gives enterprises the place to enforce a decision on each action an agent takes. Bedrock Data evaluates that action based on what the data in the request is and what company policy allows, and returns the decision. Together, NVIDIA OpenShell and Bedrock Data let enterprises scale their use of agents while company policy controls what agents are allowed to do with sensitive enterprise data."

Permissions govern which resources and operations an agent may use, but they are not data-aware: within a permitted operation, they cannot tell sensitive data from harmless data. An agent can hold exactly the permissions its job requires and still create exposure: to complete an interim computation, it writes sensitive data to a shared scratch table its job requires it to use. Every permission was correctly scoped, but none could express which data may be written there; as a result, data that was tightly restricted is now readable by everyone with access to the table. A sandbox can stop the write, but deciding when to stop it requires knowing what the data is, where it came from, what it has been combined with and where it can go. Bedrock Data evaluates an agent's action against data policy and tells OpenShell when the data involved is or isn't cleared for the destination; OpenShell can block the action accordingly.