Understanding MCP Policies

Overview

BlueRock Secure MCP Server enforces runtime security controls through configurable MCP Protection Policies.

Policies are evaluated at runtime for both MCP client requests and MCP server responses. They operate as an enforcement layer and do not require infrastructure redeployment.

When remediation is enabled (remediate: true), the MCP client or server program execution is terminated upon a policy violation. When remediation is disabled (remediate: false), BlueRock Secure MCP Server operates in observe mode and logs policy violations while allowing execution to continue.

MCP policies can be applied to MCP servers implemented in:

• Python MCP servers • JavaScript / TypeScript MCP servers

Policies regulate:

  • MCP client-server connections

  • Transport protocols (HTTP, SSE, stdio)

  • Authentication requirements

  • Tool execution behavior

  • Resource access controls

  • Prompt invocation restrictions

  • Pattern-based detection in arguments and responses


Policy Categories

MCP policies are logically grouped based on the type of control they enforce.

1. Connection Control Policies

Control how MCP clients connect to MCP servers.

Capabilities include:

  • Restricting stdio, HTTP, or SSE transports

  • Enforcing authentication for HTTP/SSE connections

  • Configuring exception lists for specific MCP server programs or URLs


2. Tool Execution Policies

Control how MCP clients invoke tools exposed by MCP servers.

Capabilities include:

  • Restricting execution of specific tools

  • Detecting forbidden patterns in tool arguments

  • Inspecting tool responses for unsafe content


3. Resource Access Policies

Control access to MCP-exposed resources.

Capabilities include:

  • Denying specific resource URIs

  • Applying server-specific resource rules


4. Prompt Execution Policies

Control prompt execution behavior within MCP interactions.

Capabilities include:

  • Denying specific prompts from specific MCP server

  • Detecting unsafe prompt arguments

  • Inspecting prompt responses for restricted patterns


5. Built-in MCP request/response message audit policies

Provide predefined detection for common risky patterns such as:

  • Dangerous shell commands

  • Privilege escalation attempts

  • Sensitive file paths

  • Suspicious network commands

  • Potential data exfiltration behavior

These apply across tool calls, prompt execution, and resource interactions.


MCP Protection Policies – Use Cases

This section demonstrates runtime enforcement behavior for MCP client-server interactions in 26.08.0 Release.

Prerequisites:

This section assumes the following setup is completed:

• Create an MCP project directory • Create a Python virtual environment • Create FastMCP client and FastMCP server programs • Install the fastmcp package • Install the bluepython package • Load the bluepython package before executing the client or server program


MCP Connection Control


Deny Client Connection to stdio Server

Policy Configuration

Running MCP Client with MCP stdio Server

Expected Behavior

Policy Configuration
Result

remediate: false

Connection allowed, violation logged (WARN)

remediate: true

Connection blocked, violation logged as ERROR, program terminated

OTEL Violation (Observe)

OTEL Violation (Enforce)


Deny stdio with Exception

Policy Configuration

Running MCP Client with MCP stdio Server

Expected Behavior

Matching command in exception list is allowed. No policy violation event is generated. A source event for the stdio connection is emitted.

Source Event


Deny Client Connection to HTTP Server

Policy Configuration

Running MCP Client with MCP HTTP Server

Expected Behavior

Mode
Result

remediate: false

Connection succeeds, WARN logged

remediate: true

Connection blocked

OTEL Violation (Observe)

OTEL Violation (Enforce)


Deny HTTP Server with Exception

Policy Configuration

Running MCP Client with MCP HTTP Server

Expected Behavior

HTTP connections blocked by default. URL in exception list is allowed. No violation event generated.

Example Source Event


Deny any mcp server connection without authentication

Policy Configuration

Running MCP Client with MCP HTTP Server

Expected Behavior

Mode
Result

remediate: false

Unauthenticated connection allowed, WARN logged

remediate: true

Unauthenticated connection blocked

OTEL Violation (Observe)

OTEL Violation (Enforce)


MCP Tool Execution Control


Tool Overwrite - Denied tools are excluded from the tool/list

Policy Configuration

Running MCP Client with MCP HTTP Server

Expected Behavior

Mode
Result

remediate: false

Tool is listed and violation logged

remediate: true

Tool is excluded from the list of tools.

OTEL Violation (Observe)

OTEL Violation (Enforce)


Deny Specific Tool from Specific MCP Server

Policy Configuration

Running MCP Client with MCP HTTP Server

Expected Behavior

Mode
Result

remediate: false

Tool executes, violation logged

remediate: true

Tool call blocked

OTEL Violation (Observe)

OTEL Violation (Enforce)


Tool Argument Pattern Detection

Policy Configuration

Running MCP Client with MCP HTTP Server

If the tool argument contains commands matching the built-in forbidden patterns (for example cat, rm, grep, etc.), the request triggers the built-in MCP audit policy.

Expected Behavior

Mode
Result

remediate: false

Tool executes, WARN logged

remediate: true

Tool blocked

OTEL Violation (Observe)

OTEL Violation (Enforce)


Tool Response Pattern Detection

Policy Configuration

Running MCP Client with MCP HTTP Server

Expected Behavior

If tool output contains forbidden patterns:

Mode
Result

remediate: false

Response allowed, violation logged

remediate: true

Response blocked

OTEL Violation (Observe)

OTEL Violation (Enforce)


MCP Resource Access Control


Deny Specific Resource

Policy Configuration

Running MCP Client with MCP HTTP Server

Run the MCP client with a resource URI that matches the deny_list.

Expected Behavior

Mode
Result

remediate: false

Resource allowed, violation logged

remediate: true

Resource access denied

OTEL Violation (Observe)

OTEL Violation (Enforce)


MCP Prompt Execution Control


Deny Specific Prompt

Policy Configuration

Running MCP Client with MCP HTTP Server

Expected Behavior

Mode
Result

remediate: false

Prompt execution proceeds, but a violation event is generated.

remediate: true

Prompt invocation is blocked and the MCP client execution is terminated.

OTEL Violation (Observe)

OTEL Violation (Enforce)


Prompt Argument Pattern Detection

Policy Configuration

Running MCP Client with MCP HTTP Server

Expected Behavior

Mode
Result

remediate: false

Prompt execution proceeds, but a violation event is generated.

remediate: true

Prompt invocation is blocked and the MCP client execution is terminated.

OTEL Violation (Observe)

OTEL Violation (Enforce)


Prompt Response Pattern Detection

Policy Configuration

Running MCP Client with MCP HTTP Server

Expected Behavior

If response contains nc, curl, etc:

Mode
Result

remediate: false

Response allowed, violation logged

remediate: true

Response blocked

OTEL Violation (Observe)

OTEL Violation (Enforce)


Message Size Violation

Policy Configuration


Running MCP Client with MCP HTTP Server


Expected Behavior

Mode
Behavior

remediate: false

Request proceeds; violation logged (WARN)

remediate: true

Request proceeds; violation logged (WARN)

OTEL Violation (Observe)


Summary

Mode
Behavior

remediate: false

Observe mode – Action allowed, WARN logged

remediate: true

Enforce mode – Action blocked, ERROR logged

All violations are emitted as OTEL events and sent to external collector such as AWS CloudWatch based on the configurations.

  • event_name: mcp_policy_violation

  • severity_text: WARN or ERROR

  • type:

    • "log" (observe events)

    • "remediation" (enforce events)

  • remediation_kind: "block" (enforce mode only)


Last updated