BlueRock 26.08.0

Purpose of this Document

This is a living document and is updated with each release.

This document provides information about the features, enhancements, and known issues in the BlueRock Release 26.08.0. It applies to both new deployments and upgrades from earlier versions.

This focuses on expanding observability capabilities for AI agent frameworks, improving container runtime support, and enhancing the overall developer experience for MCP-based applications. These updates enable better visibility into runtime behavior and simplify deployment and execution of agentic workloads and improve the workflow for running MCP applications within the BlueRock instance.

---

Revision History

Release Version	Release Date
V26.08.0

---

Supported Platforms / Distributions

AWS

BlueRock Release 26.08.0 supports deployment on AWS EC2 instances.

Following Linux distributions are supported on EC2 and EKS deployments:

---

Supported Linux Distributions

Linux Distribution	OS Version	Kernel Version
Amazon Linux 2023	2023.10.20260120	6.12.63
Bottlerocket	v1.54.0	6.12.63

---

GCP

BlueRock Release 26.08.0 supports deployment on GCP environments.

Supported Linux Distributions

Linux Distribution	OS Version	Kernel Version
Ubuntu	24.04	NA

---

New Features and Enhancements

This release introduces expanded platform support with GCP compatibility using Ubuntu distribution, along with enhanced observability and policy enforcement across AI agent workflows. It includes new sensors for JavaScript/TypeScript, LangChain, CrewAI, A2A, and LLM model interactions, as well as improvements to the BlueRock Sandbox with container runtime support and simplified execution.

---

[BC-3044] New JS Sensor enables runtime observability and policy enforcement for JavaScript/TypeScript MCP applications.

Introduces a Node.js sensor that enables BlueRock observability for JavaScript and TypeScript MCP client and server applications with support for:

• Monitoring MCP session, tool/resource and prompt usage

• Enforcing protection policies to guard against unsafe executions

• Capturing runtime execution events

Sample policy configuration:

"js_sensor": {
  "mcp": {
    "enable": true,
    "remediate": false
  }
}

For all sensors, the feature is enabled for observability in the default policy loaded on to the system (applicable for both Free edition and Full edition).

---

[BC-3665] LangChain sensor enables observability for LangChain Agent Applications

Introduces a LangChain sensor that enables observability for AI Agent applications built using Python LangChain SDK.

Policy Configuration to enable sensor:

     {
  "langchain": {
    "enable": true
  }
}

---

[BC-3266] CrewAI sensor enables observability for CrewAI agent workflows

Introduces a CrewAI sensor that enables observability for AI agent workflows built using the CrewAI framework.

Policy Configuration to enable sensor:

{
  "crewai": {
    "enable": true
  }
}

---

[BC-3188] Enhanced version of BlueRock Sandbox with container runtime support

Introduces an enhanced version of BlueRock Sandbox with simplified command arguments and improved support for running AI agentic applications and MCP server applications using container images.

Key enhancements include:

• Simplified command arguments for improved ease of use

• Automatic mounting of required host volumes in read-only mode

• Support for passing required environment variables during sandbox execution

• Option to enable or disable namespace isolation

• Support for wildcard characters in policy configuration (exec)

Container Support:

• Enables deployment and execution of AI agentic applications and MCP server applications as Docker containers

---

[BA-2762, BA-2763, BA-2764] A2A sensor enables observability and policy enforcement for Python-based A2A applications

Introduces an A2A sensor for Python applications built using SDKs that support the Agent-to-Agent (A2A) protocol.

The sensor provides observability into agent-to-agent communication by capturing runtime events such as agent card requests and server interactions during A2A workflows.

It also enables policy enforcement on agent capabilities advertised via Agent Cards, with built-in protections to detect unsafe patterns in:

• Agent card name

• Agent card description

• Agent skill descriptions

---

[BA-2796, BA-2797, BA-2798, BA-2799] LLM Model sensor enables observability for LLM interactions in Python AI Agent applications

Introduces an LLM Model sensor that provides observability into interactions with Large Language Models (LLMs) across supported Python SDKs, including OpenAI and GoogleAI.

The sensor captures runtime events for model invocations, including request payloads (prompts and parameters) and model responses, providing visibility into application-level interactions with LLMs.

This enables traceability of model usage within AI agent workflows, including prompt execution, tool invocations, and model responses with full traceability.

---

Resolved Issues

No resolved issues in this release.

---

Known Issues

The following Known issues are present in this release.

[BC-3256] Client-side tool deny policies do not remove denied tools during discovery with JavaScript MCP sensors

When client-side tool deny policies are configured, denied tools are not removed from the tool discovery (tool/list) response when using JavaScript or TypeScript MCP sensors.

As a result, denied tools continue to appear in the discovered tool list, and associated remediation actions and violation events are not triggered.

This issue affects JavaScript and TypeScript MCP servers. Python MCP sensors enforce deny policies correctly.

Workaround: Recommended to use Server side tool filter policy.

Policy Example:

"js_sensor": {
  "mcp": {
    "enable": true,
    "remediate": true,
    "server": {
      "enable": true,
      "remediate": true,
      "tools": {
        "FileAgentServer": {
          "deny_list": ["remove_file"]
        }
      }
    }
  }
}

---

[BC-3252, BC-3249, BC-3282] Limitations in BlueRock Sandbox Networking and Execution

The following limitations are observed when using the BlueRock Sandbox environment.

[BC-3252] Restart Failure After Network Policy Update

After updating a network policy, restarting an Agent or MCP application container in the BlueRock Sandbox may fail to start.

Workaround: Restart the BlueRock control plane service, then restart the application container.

sudo systemctl restart uc-docker.service

---

[BC-3249] CLI Argument Syntax Limitation

The BlueRock Sandbox CLI does not support space-separated syntax for the --network-config argument.

• Not Supported:

--network-config <config_name>

• Supported:

--network-config=<config_name>

Workaround: Use the equals (=) syntax when passing the --network-config argument.

---

[BC-3282] Host Binding Does Not Expose Application Ports

When using network configuration with ingress rules in the Sandbox, application ports are not exposed to the host layer as expected.

• Applications are not accessible via host IP and port

• External connections fail

When the sandbox is run without network configuration ingress rules, ports are correctly exposed.

Workaround: Avoid configuring ingress rules. Run the sandbox without network configuration.

---

[BC-3260] Python application execution hangs after repeated Ctrl+C termination when using bluepython module

When running Python applications using the BlueRock Python sensor (bluepython module), repeated termination of the application using Ctrl+C (typically after 2–3 runs) may result in subsequent executions hanging. This problem is a rare occurrence.

Workaround:

Restart the BlueRock Control Plane service:

sudo systemctl restart uc-docker.service

---