Skip to content
    Gaincontrol

    Pilots open · SDK + control plane · 4-week program

    Every agent call, authorized before it runs.

    Zino gives each AI agent its own verifiable identity and checks every tool, network and MCP call against policy before it executes. Denied calls never run, credentials never sit in the agent, and every decision lands in a tamper-evident log.

    policy: payments-prod · v12 (signed), enforcing: 4 allowed · 2 denied · every decision logged

    • refund-agent → payments.refund($48): allow, within grant (#a91f)
    • refund-agent → payments.refund($48,000): deny, over $500 limit (#e49c)
    • support-bot → mcp:github.create_pr: allow, approved MCP tool (#2019)
    • refund-agent → POST api.stripe.com: allow, within grant (#5b96)
    • triage-agent → GET 185.220.101.4: deny, destination not on allowlist (#9713)
    • etl-agent → db.orders.read: allow, within grant (#d290)

    Follow one hijacked request through Zino.

    An instruction hidden in a support ticket tells refund-agent to send $48,000 to an outside account. Here is each place Zino stops it, and the evidence each stop leaves for your auditor.

    1. Who is asking?

      Zino knows the request comes from refund-agent itself, through a verifiable identity of its own, not a borrowed human token. Every decision after this is tied to that one agent.

      Verified: refund-agent

      StopsBorrowed tokens · Impersonation · Shared service accounts

      Evidence for

      Every decision names a verified agent, not a shared key.

      • SOC 2
      • ISO 27001
      • NIST AI RMF
    2. Is it allowed to do this?

      Before the call runs, Zino checks it against a signed policy. Refunds are capped at $500 per call, so this one is denied, whatever the ticket told the agent to do. If the policy itself were tampered with, Zino would refuse to run rather than fail open.

      Denied: over the $500 limit

      StopsPrompt injection · Excessive agency · Tampered policy

      Evidence for

      The signed policy version and the reason behind every denial.

      • OWASP LLM Top 10
      • SOC 2
      • ISO 27001
      • EU AI Act
    3. What if it tries another way?

      The agent tries to reach the payment provider directly, then through an unvetted MCP tool. Zino checks network and MCP calls too, so destinations off the allowlist and unapproved tools are blocked.

      Blocked: not on the allowlist

      StopsData exfiltration · MCP tool poisoning · Policy bypass

      Evidence for

      Every blocked destination and unapproved tool, by agent.

      • OWASP LLM Top 10
      • EU AI Act
      • NIST AI RMF
    4. Can it steal a key instead?

      There is no long-lived key in the agent to steal. Zino brokers credentials from your existing vault and injects them per approved call, so a hijacked agent has nothing to take.

      Nothing to steal

      StopsSecret sprawl · Stolen API keys

      Evidence for

      Credentials issued per approved call, and no standing secrets to account for.

      • SOC 2
      • ISO 27001
    5. Can you stop it, and prove what happened?

      Your team pauses the agent in seconds with the kill switch, with no redeploy. Every attempt sits in a tamper-evident, hash-chained log that your auditor can check independently.

      Paused · every attempt logged

      StopsRogue agents · Missing audit trail

      Evidence for

      A hash-chained record of every attempt, and when access was paused.

      • EU AI Act
      • NIST AI RMF
      • SOC 2
      • ISO 27001
    6. No money moved. The agent was contained, and every attempt is on record.

    Evidence mapping, not a certification claim. Your auditor decides scope.

    One control plane. Every workload enforced in-process.

    The Policy Server is the single signing authority for the fleet. The SDK runs inside each workload — no sidecar, no network hop for policy evaluation — and gates every REST, Shell, and MCP call at the trust boundary before it leaves.

    Enterprise Vault

    your existing secrets store

    brokers secrets

    Policy Server

    Control plane · one authority

    ↓ identity + signed policy

    ↑ hash-chained audit

    Workload

    Zino SDK

    in-process · no sidecar · credential injected per call, never held

    Trust boundary

    REST / HTTP

    checked here

    Shell

    checked here

    MCP

    checked here

    … and every other workload in the fleet

    Governs what you already run.

    Zino policy-controls identity and access for the systems your agents already touch. No migration: one SDK call instruments each agent.

    Integrations, not customers: these are the systems Zino can govern.

    GitHub
    GitLab
    Jira
    Okta
    SAP OData
    Stripe
    Slack
    Salesforce
    Azure DevOps
    ServiceNow
    Microsoft Dynamics 365

    Frequently asked questions

    Zero-trust agent governance means no AI agent is trusted by default. Every agent gets a verified identity, every action is denied unless explicitly authorized at runtime, every decision is written to a cryptographic audit trail, and access can be suspended or revoked instantly if an agent misbehaves.

    Zino grants AI agents only the permissions required for their current task, automatically revoking access when the task completes. Permissions are scoped dynamically rather than assigned statically at deployment time.

    Yes. Zino produces signed, hash-chained access logs that are structured for SOC 2 Type II, ISO 27001, and other regulatory frameworks. Auditors can query agent access history without requiring access to production systems.
    All posts
    • · 3 min read

      Authenticating the agent isn't authenticating the action

      "Which agent is this" and "is this action allowed right now" are different questions. Most systems built for agent authentication only ever answer the first one, and treat that as settling the second.

    • · 3 min read

      What a tamper-evident audit trail actually has to prove

      Most "audit logs" built for AI agents are mutable, editable by whoever already has database access, and easy to doctor without leaving a trace. Tamper-evidence is a specific, checkable property, not an adjective a vendor can just claim.

    • · 3 min read

      MCP poisoning: when the server you trusted isn't the one answering today

      An MCP allowlist only ever verified a name at one point in time. If whatever answers to that name changes — a tampered registry entry, a lookalike server, a compromised operator — the allowlist entry still matches, and nothing says so.

    Zino governs identity and authorization. Quin finds what to govern, and Drona, in private development, will orchestrate the workflow around it.

    Govern your agents from day one.

    Zino is built for security and platform teams who know that zero-trust agent governance can't wait for a breach to prove its value.

    Book a demoPilot · 4 weeks · 1 sponsor · 1 fleet · no infrastructure liftSee the pilot plan →