Cloud & homegrown agents

Runtime protection,inside your cloud.

Agents built on AWS Bedrock, Azure OpenAI and Vertex AI call models, MCP servers and internal APIs. Zenstra decides those calls at a gateway in your VPC; payloads never leave.

The risk

Your agents reach
your most sensitive systems.

Homegrown agents are built to do real work: read accounts, update cases, move money. That makes every call they make a decision worth checking.

  • Over-scoped credentialsModel and API keys shared across agents and environments.
  • Tools everywhereMCP servers and internal APIs added faster than reviews.
  • Data residencyPrompts and payloads that must not leave your cloud.
A real pathA crafted request asks a loan agent to move funds.
  1. 1Bedrock agentA loan-servicing agent built on AWS Bedrock runs inside your VPC.
  2. 2grantIt asks for access to the loan system and receives a grant for 15 minutes.
  3. 3requestA crafted request asks it to transfer funds between accounts.
  4. Blockaccounts.transfer is not in the grant. Stopped at the gateway; payloads stay in your VPC.mcp core-banking · accounts.transfer
Industry references
  • ASI03Identity and Privilege AbuseOWASP Top 10 for Agentic Apps
  • LLM06:2025Excessive AgencyOWASP Top 10 for LLMs
  • AML.T0053AI Agent Tool InvocationMITRE ATLAS
Policy as code

How it is
stopped.

Per-agent identity and short-lived grants for the models and tools each agent uses, decided at a gateway in your own cloud.

  • Decided inline
  • Fails closed
  • Versioned with every decision
allowbedrock.invokefor loan-doc-classifier · expires 15mdenymcp.callwhen tool not in grantholdpayments.*for human approvaldenynet.egressto unapproved destinations
How Zenstra handles it

A decision point
in your own environment.

The gateway runs where your agents run, in front of every model, MCP server and API.

Deploy

In your VPC, your way

Deploy the gateway with Helm or Terraform in AWS, Azure or Google Cloud, in your VPC or Kubernetes cluster.

  • Helm or Terraform
  • AWS, Azure or Google Cloud
  • No agent rewrite
Scope

Per-agent identity and grants

Each agent gets its own identity and short-lived grants for the models and tools it uses.

  • Scoped credentials for Bedrock, Azure OpenAI and Vertex AI
  • Per-call and per-session grants
  • Revocation in seconds
Decide

Fail-closed, fast

Every call is decided inline, before it reaches the tool. If Zenstra is unreachable, actions are held.

  • Inline decisions
  • Fails closed
  • Every decision on the record
What it protects

Every call
your agents make.

  • Model calls

    Bedrock, Azure OpenAI and Vertex AI, per-agent credentials.

  • MCP servers

    Each tool call checked against the agent’s grant.

  • Internal APIs

    Core, LOS and CRM connectors as scoped grants.

  • Payments

    Held for a person to approve, by policy.

  • Outbound data

    Approved destinations only.

  • Agent identity

    Attested before actions run.

FAQ

Common questions.

Does the gateway add latency?

Decisions are made at 18 ms at p95, inline with the call.

What reaches the Zenstra control plane?

Only decisions and redacted evidence. Prompts and payloads stay in your environment.

Which clouds are supported?

AWS, Azure and Google Cloud, deployed in your VPC or Kubernetes cluster.