Data & credential leakage

Member data and secretsstay where they belong.

Agents move data as part of their work, and that is how it leaks. Zenstra decides each email, export and model call before it runs.

The risk

Leaks look like
ordinary actions.

Sending an email, exporting a file and calling a model are normal actions. Whether they leak depends on task, data and destination.

  • Member dataAccount references sent outside the institution.
  • CredentialsKeys and secrets read or pasted into AI tools.
  • Payloads to modelsSensitive prompts sent to unapproved endpoints.
A real pathA member’s file heads for an outside inbox.
  1. 1loan-doc-classifierThe agent reads a member’s loan application to sort its documents.
  2. 2email.sendA step in the flow tries to email the file to an address outside your domains.
  3. 3POL-114Policy checks the data, the destination and the task before anything leaves.
  4. BlockMember data egress blocked, and recorded with the policy that decided it.POL-114 · email.send · external
Industry references
  • LLM02:2025Sensitive Information DisclosureOWASP Top 10 for LLMs
  • AML.T0086Exfiltration via AI Agent Tool InvocationMITRE ATLAS
  • AML.T0057LLM Data LeakageMITRE ATLAS
Policy as code

How it is
stopped.

Whether an action leaks depends on the data, the destination and the task. Zenstra decides on all three, before it runs.

  • Decided inline
  • Fails closed
  • Versioned with every decision
denyemail.sendwith member data · to external domainsdenymodel.callwith secrets in payloadholdexport.*over 100 recordsallowcrm.readfor member in task
How Zenstra handles it

Task, data and destination,
checked together.

Decided inline, before the action runs. If Zenstra is unreachable, the action is held.

Scope

Only the data the task needs

Reads are limited to the member and the records in the task, through scoped, expiring grants.

  • Read the requesting member only
  • Per-call grants
  • Signed writes only
Stop

Egress decided before it happens

Outbound actions are checked against approved destinations and the declared task, and blocked when they do not fit.

  • Approved destinations for email and export
  • Intent check on outbound actions
  • Secrets access contained on devices
Prove

What left, and what never did

The record shows every outbound decision, and the security review pack lists what reaches the control plane.

  • Every egress decision recorded
  • Payloads stay in your environment
  • Redacted evidence only
Leak paths

Where data
escapes.

  • Outbound email

    Agents asked, or tricked, into emailing records.

  • Exports and reports

    Files generated with more data than requested.

  • Model calls

    Sensitive prompts sent to unapproved models.

  • Secrets on devices

    Coding agents reading cloud keys and tokens.

  • Pasted into AI

    Staff pasting secrets and member data into chat tools.

  • Agent hand-offs

    Data passed to another agent outside the task.

FAQ

Common questions.

Is this a DLP product?

No. DLP inspects content; Zenstra decides agent actions with the context of the task, the agent and the destination. They work well together.

Where is data evaluated?

Inside your environment. Only decisions and redacted evidence reach the Zenstra control plane.

Does it cover secrets on laptops?

Yes. Zenstra AI Discovery contains coding agents that try to read secrets they were never given.