Insights

What is an MCP server, and how do you secure its connections?

19 November 2026

Review MCP server security with practical checks for provenance, authentication, tool permissions and logging, while retaining human approval.

ThreatLocker has published an explainer on MCP servers and their security risks. It highlights an important issue for businesses connecting AI agents to operational systems: a useful integration can also create a new route to sensitive information and privileged actions.

For Australian IT leaders, MCP server security should be a connection approval process, not simply a software installation check. Even when agents run in a private Australian-hosted environment with human approval retained, every connected server, credential and downstream system needs its own controls.

What is an MCP server?

Model Context Protocol, or MCP, is a standard way for AI applications to connect to tools and information sources. An agent’s MCP client communicates with a server, which exposes selected capabilities from an application, database or other service.

For example, a ticketing integration might provide separate tools to search tickets, retrieve a ticket and update its status. An agent can request one of those operations through the connection rather than requiring a different integration pattern for every application.

An MCP server is not necessarily a separate physical machine. It may be a local process, a service in your private environment or a remote endpoint operated by a supplier. That distinction affects how it is authenticated, isolated and maintained.

MCP standardises communication. It does not establish that a server is trustworthy, that its permissions are appropriate or that an agent’s proposed action should proceed.

Why MCP server security needs separate controls

ThreatLocker’s explainer emphasises that access to connected tools can increase the consequences of a compromised or manipulated agent. The practical concern is the chain from incoming content to a tool call and then a business action.

A document containing malicious instructions might influence an agent to retrieve unrelated records. A vulnerable server could expose credentials. An overly broad service account could allow a legitimate tool to change records outside the intended workflow.

The key risks include:

  • Untrusted software: a package, dependency or update introduces malicious behaviour or a vulnerability.
  • Excessive access: a narrow task inherits broad database, file or application permissions.
  • Prompt injection: instructions embedded in retrieved content attempt to redirect the agent.
  • Cross-user exposure: a shared server fails to keep users, workflows or customers separated.
  • Missing evidence: logs show a successful connection but not the resulting data access or change.

Apply Zero Trust principles at each boundary: verify the connection, restrict its scope and evaluate each requested action. An instruction telling the agent to “be careful” is not an access control.

1. Verify server provenance before installation

Start by establishing exactly what you are approving. A familiar product name in a repository or directory listing is not proof that the application vendor publishes or supports that server.

Record the publisher, source repository or distribution channel, deployment location, version and internal owner. Confirm whether the server is vendor-maintained, community-maintained or custom-built, and who will respond when a vulnerability is reported.

Before connecting production data:

  • Review the requested installation commands, dependencies and runtime privileges.
  • Use a controlled package source and pin the approved release where practical.
  • Check available signatures or integrity hashes against a trusted publisher channel.
  • Inspect exposed tools, descriptions and configuration options.
  • Test updates outside production and review changes to tools or permissions.

For locally executed servers, avoid running them with workstation administrator rights or unrestricted access to shared folders. For remote services, verify the endpoint and the supplier’s maintenance arrangements.

Evidence to retain: an approved server inventory entry, version record, technical review and named patching owner.

2. Check authentication at both ends

There are two distinct questions: how does the agent authenticate to the MCP server, and how does that server authenticate to the business application?

A secure front door does not compensate for a shared administrator credential behind it. Prefer workload-specific identities and narrowly scoped credentials. Where user-specific access is required, confirm that the downstream system receives and enforces the correct user context.

For HTTP-based deployments using OAuth, check token validation, intended audience and scope enforcement. The server should not blindly accept an incoming token and forward it to another service. Local process-based deployments need appropriate operating-system isolation and protected credential handling instead; they do not necessarily use the same authentication flow.

Store secrets outside prompts and source code. Set rotation and revocation procedures, and test what happens when an identity is disabled.

Evidence to retain: an identity map covering client, server and downstream application, plus a successful revocation test. Our AI agent identity security checklist expands on account ownership and approval responsibilities.

3. Limit tools, records and human-approved actions

Approve capabilities individually. “Access to finance” is too vague; “read approved purchase orders for one business entity” is a permission you can assess and test.

Begin with read-only tools wherever possible. Separate retrieval, draft creation and final submission. Avoid exposing general-purpose shell commands, unrestricted SQL or broad file access when a narrow business tool will do.

For an AP/AR Automation Agent, the permitted sequence might be to read an invoice, compare it with authorised records and prepare a draft entry. Posting an entry or changing supplier bank details should remain outside that permission set unless separately authorised and technically controlled.

Human approval must be enforced outside the model’s own instructions. The approval screen should identify the exact action, affected records, destination and material changes. Bind approval to that specific action; if the parameters change, require fresh approval.

Retain human approval for consequential changes, external communications and sensitive disclosures. Pre-approved, tightly bounded retrieval can support preparation without giving the agent authority to approve its own outcomes.

Evidence to retain: a tool-permission matrix and tests showing that forbidden records, actions and unapproved writes are rejected.

4. Confirm private hosting does not hide external data flows

Private hosting gives you control over the agent environment, but a connected tool can still send information outside it. Review MCP server security alongside the complete data path, including model endpoints, telemetry, support access and backups.

For agents in a private Australian-hosted environment, document every permitted destination. Restrict outbound connectivity to approved services and keep development credentials separate from production credentials.

Ask what information each tool transmits, where responses are processed and whether logs capture business content. Check that the hosting and processing arrangements match your contractual and privacy requirements rather than assuming the server’s location settles the question.

Treat returned documents and tool descriptions as potentially untrusted inputs. They should not be able to grant new permissions or bypass approval gates.

Evidence to retain: a data-flow diagram and tested outbound access rules. Use our data sovereignty buyer checklist to assess more than the hosting address.

5. Log the action chain and test shutdown

Useful logging should connect the initiating user or workflow to the agent, MCP server, selected tool, approval decision and downstream result. A record saying “tool executed” is not enough to investigate an unexpected change.

Capture timestamps, correlation identifiers, server versions, relevant record identifiers, outcomes and policy denials. Record parameters only to the extent needed for investigation, with sensitive fields redacted. Never log access tokens or secrets.

Restrict log access and protect records against alteration. Define retention based on business, privacy and incident-response requirements. Alert on unusual exports, repeated authorisation failures, new destinations and unexpected write activity.

Before launch, test a malicious instruction in a sample document, an attempted cross-user lookup and an unapproved write. Confirm that controls block the actions and that investigators can reconstruct what happened.

Finally, exercise a shutdown procedure: disable the connection, revoke credentials and stop pending jobs. Assign an owner who can act without waiting for a supplier discussion.

Evidence to retain: a traceable test transaction, blocked-action logs and a documented shutdown test.

Make connection approval a release gate

Approve an MCP connection only when its owner can demonstrate provenance, scoped authentication, restricted tools, enforceable human approval, understood data flows and useful logs. Repeat the review when the server, tools, credentials or destinations materially change.

Tech Engine can help review your cyber security and implement managed AI agents that integrate with existing systems in a private Australian-hosted environment, with human approval retained. Discuss your proposed connection at aiaas.au, call 1300 088 324 or email sales@techengine.au.

Want this applied to your business?

Request an AI Blueprint and we will map the processes worth automating first.

Request an AI Blueprint