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.
refund-agent · Refund $48,000 to an outside account
Prompted by an instruction hidden in ticket #8812
- IdentityVerified: refund-agent
- PolicyDenied: over the $500 limit
- Network & MCPBlocked: not on the allowlist
- CredentialsNothing to steal
- Containment & auditPaused · every attempt logged
Evidence collected so far
- SOC 24 checkpoints
- ISO 270014 checkpoints
- NIST AI RMF3 checkpoints
- EU AI Act3 checkpoints
- OWASP LLM Top 102 checkpoints
No money moved. The agent was contained, and every attempt is on record.
1 / 5 · Identity
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 / 5 · Policy
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 / 5 · Network & MCP
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 / 5 · Credentials
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 / 5 · Containment & audit
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
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
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
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.
Frequently asked questions
Related reading
All posts· 3 min read
· 3 min read
· 3 min read
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 →