Runtime policy enforcement

Every action decidedbefore it runs.

One policy engine checks every agent action against identity, task and policy, inline at 18 ms p95. Allowed, held for a person, or blocked.

The problem

Agents act faster
than your controls.

Alerts arrive after the money moved and the data left. For agents, a control that reports later is a control that was too late.

  • After the factLogs and alerts describe what an agent already did.
  • Policy on paperAcceptable-use policies written as PDFs that no agent can read.
  • One rulebook per vendorEvery platform has its own switches, and none of them agree.
A real pathAn agent asks for more data than its task needs.
  1. 1sar-prep-agentThe agent is asked to prepare a suspicious activity report for one member.
  2. 2planIt plans core.txn.query across 90 days of transactions.
  3. 3ZenstraThe policy engine checks identity, task and policy inline, before the query runs.
  4. HoldHeld for a person to approve, before anything ran.sar-prep-agent · core.txn.query (90d)
Industry references
  • LLM06:2025Excessive AgencyOWASP Top 10 for LLMs
  • ASI02Tool Misuse and ExploitationOWASP Top 10 for Agentic Apps
  • AML.T0053AI Agent Tool InvocationMITRE ATLAS
How Zenstra handles it

Your policy,
enforced as code.

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

Check

Identity, task and policy together

Each action is checked against who the agent is, what it was asked to do and what your policy allows.

  • Inline decision, 18 ms at p95
  • Intent checked against the declared task
  • Agent-to-agent hand-offs checked against the original task
Decide

Allow, hold or block

Routine actions pass. Sensitive ones wait for a person. Anything outside policy is stopped, and the agent continues safely where it can.

  • Hold payments and production changes for approval
  • Block data leaving to unapproved destinations
  • Fails closed: if Zenstra is unreachable, actions are held
Record

Why, for every decision

Every decision lands on the record with the policy version that made it, so you can show what was allowed and why.

  • Policy version on every entry
  • Approvals recorded with who and when
  • Tamper-evident, chained record
Policy as code

The policy
behind it.

One engine and one rulebook for every agent. Routine actions pass, sensitive ones wait for a person, anything outside policy is stopped.

  • Decided inline
  • Fails closed
  • Versioned with every decision
denytool.callwhen outside(task.scope)holdpayments.*for human approvalholdcore.txn.querywhen range > 30dallowcrm.readfor member in task
What policies cover

Write the rule once.
It holds everywhere.

  • Task scope

    deny tool.call when outside(task.scope).

  • Human approval

    hold payments.* for human approval.

  • Member data

    allow crm.read for the member in the task only.

  • Outbound data

    Block email and exports to unapproved destinations.

  • Production access

    Step-up before shell or deploy actions in production.

  • Acceptable use

    Your AI acceptable-use policy as enforced code, not a PDF.

FAQ

Common questions.

Does enforcement slow agents down?

Decisions are made inline at 18 ms at p95, well inside the time a model call takes.

What happens if Zenstra is unavailable?

It fails closed: actions are held rather than allowed, so an outage never becomes an open door.

Do we have to rewrite our agents?

No. Enforcement sits in front of the agents you have, through the SDK, the gateway in your cloud or the AI Discovery app.