Bedrock Data ads

Least privilege is not least risk. Today we released Bedrock Data Agent DLP for NVIDIA OpenShell. OpenShell decides whether an agent may perform an action. Agent DLP decides whether the data in that action may go where it is headed. An agent with the right permissions can still move intellectual property somewhere policy says it cannot go, because permissions are not data-aware. Agent DLP sits at OpenShell's supervisor middleware, identifies what the request contains, and blocks it before it leaves the sandbox. I wrote up why least privilege stops short and how the integration works. Link in the comments. We're excited to continue building alongside Ali Golshan, Alex Watson and the NVIDIA OpenShell team.

On Oct 20 I’m speaking at the IANS New York Forum on least privilege != least risk. Here’s a preview of my talk track. An agent with exactly the permissions its job requires, not one scope more, can still leak your sensitive data. Consider a root-cause agent scoped to least privilege: it can read code and open issues, nothing more. It traces a backend failure to an open-source library, writes a reproduction and files a public GitHub issue, which is what a good engineer would do. The repro carries the company’s pricing logic with it. Every permission in that chain was correct and a more narrow scope would have stopped it. Permissions answer "can this agent perform action X?" They don’t answer "is this data allowed for this action?" That’s the gap least privilege leaves open. This gap shows up three ways: 1. Toxic combinations: Two sources that are fine on their own become sensitive the moment an agent joins them. 2. Data drift: The data behind a permission changes after the permission was granted. The scope stays the same but the risk doesn’t. 3. Agent-to-agent handoffs: One agent passes context to the next, and the data travels with it into places nobody cleared it for. Data authorization asks whether a data payload is allowed for a given action in its current execution context. In the example above, a data authorization check would have flagged that the payload contained company intellectual property in a create_issue tool call to a non-company repository. If you’re at IANS New York, come find me after!

Your environment is cloud-first, AI-driven, and moving fast, while the tool watching it was designed for a slower, simpler estate. Bedrock Data is built for how data works now: petabyte scale, AI-native classification, and a clear view of who and what can access your data. Get your free data estate assessment.

AI moves your risk surface in real time, but a legacy DSPM platform updates on a cadence measured in quarters, so it can never keep up with what your agents are doing. Bedrock Data is serverless and continuous, with a live view of who and what can access your data. Get your free data estate assessment.

AI moved your risk surface into new territory that a last-gen tool was never built to defend, so you are outmatched before you start. Bedrock Data is AI-native and built for it: full coverage in hours, and a live view of what every AI agent and copilot can access. Get your free data estate assessment.

In his new blog, Bedrock Data CTO Pranava Adduri explains why agent sandboxes need data context and how Bedrock Data Agent DLP and NVIDIA OpenShell bring it to every outbound action. Pranava shares an example of how a root-cause agent posted its company’s pricing logic to a public GitHub issue while using only the permissions its job required. To stop that request, OpenShell checks whether the operation is permitted, then Agent DLP uses the Metadata Lake to identify what the payload contains and returns a decision that OpenShell enforces before the request leaves the sandbox. Teams can start in observe-only mode and turn on enforcement one agent at a time. Read the 5 steps Pranava recommends before deploying an agent with sensitive permissions: https://lnkd.in/g7uVQNzM Learn more about AgentDLP: https://lnkd.in/gDzFDXrn #AIGovernance #DataSecurity