On Tuesday, November 17, Virtru is launching the Virtru Data Harbor Platform in Washington, DC, and online. We'll show how enterprises and agencies are enforcing access policy on the data itself, one object and one request at a time, and why that has become the only reliable way to govern AI agents.
In 2022, generative AI was a chat window with a person deciding what to do next. Today, agents retrieve documents, query databases, and hand results to other agents thousands of times a day, with no human reviewing each request.
The policies exist, but most organizations can't enforce them. In Delinea's 2026 Identity Security Report, 99.7% of organizations said they have a formal policy on what data AI agents may access. Only 51% check agent access against it in real time, and only 19% can detect in real time when an agent reaches data outside its scope. Asking whether an agent can reach a system isn't enough anymore. The question is whether this agent may open this object for this purpose right now, and it has to be asked again on every request.
Most AI security investment goes to detecting what agents do and deciding what they're allowed to do. Identity platforms, gateways, runtime tools, and posture management are all necessary, and Virtru works alongside each of them. But they enforce by proxy, at a gateway, endpoint, app, or tenant edge. That leaves three gaps:
The Data Harbor Platform is built on the Trusted Data Format (TDF), an open standard. TDF cryptographically binds attributes and policy to each data object. When an agent requests access to a particular data object – the agent's entitlements are evaluated against that object's attributes. If the request is aligned with policy, then the key is released and the object is decrypted. If not, the object remains encrypted.
This uniquely granular approach to data security combines five properties that proxy-based tools don't: policy bound to the object, attributes tested on every request, control that works across organizational boundaries, and an open standard. It also makes policy enforcement and audit a single event. Every key request records which agent, which object, which policy, and what decision, and that record follows the object across firms.
One global financial institution evaluated a long list of AI security vendors. Each could detect agent activity or decide entitlements, but none enforced policy on the data once an agent had legitimate access. In six weeks, the bank and Virtru built a working prototype on the Platform. It ran self-managed and air-gapped, with keys and policy staying inside the bank. The bank's policy engine made the decisions, and agents could only access the data they were entitled to. Unauthorized agents were refused even where the general business rule would have allowed access, and every refusal was recorded. In the bank's words: "We can prove what data was accessed and what was refused."
Skeptics will say object-level encryption adds key management overhead, slows agents down at scale, and makes data harder to search and analyze. They’ll add that an authorized agent with a key reads the data like anyone else. That’s true, and it’s the point. A rogue agent is dangerous because it’s authorized, so the only control that can stop it is one that checks every request at the data itself.
Think about what that agent looks like. It has valid credentials, it sits inside the tenant, and more often than not -- it’s misaligned rather than malicious. Every control that governs where an identity can go has already let it in. The one checkpoint it can’t get around is the key: no key, no readable data. Virtru releases the key per agent, per object, per request, based on attributes the data carries wherever it goes. Being authorized to reach a system isn’t the same as being entitled to open a specific object, and only a control bound to the object can tell the two apart.
The practical objections have practical answers. You encrypt only the data that matters, which your classification tools already identify. A key request is small next to the work an agent does with the data. And policy comes from systems you already run, so there’s no new rulebook for AI. Whether an agent is rogue, compromised, or misaligned, it can expose only what it’s entitled to. That’s what makes it safe to give agents real data.
See the Platform in action and hear from the partners who built on it, in Washington, DC, or online.