In this series, we've explored the security landscape of AI coding agents in CI/CD environments, examining the fundamental challenges, GitHub Copilot's built-in firewall approach, and Claude Code's unrestricted architecture. Now, let's examine Google's Gemini CLI - an agent that takes a unique middle path with flexible authentication and built-in observability, yet still requires runtime security monitoring for complete visibility.

Gemini's Hybrid Approach: Flexibility with Observability

Google's run-gemini-cli GitHub Action represents a distinctive philosophy in AI coding agents. Unlike Copilot's restrictive firewall or Claude's completely open architecture, Gemini offers a hybrid approach that emphasizes flexibility in deployment while providing native observability features through Google Cloud integration.

This design allows Gemini to:

• Operate with either simple API keys or enterprise-grade Workload Identity Federation

• Send telemetry data (traces, metrics, and logs) to your Google Cloud project

• Leverage Vertex AI for enhanced capabilities when needed

• Integrate seamlessly with Google's ecosystem while maintaining GitHub-native workflows

However, this observability focuses on application-level metrics rather than runtime security. While you can monitor Gemini's performance and API usage through Google Cloud, you lack visibility into:

• Which external endpoints the agent contacts during execution

• What processes are spawned by agent-generated code

• File system modifications made during workflow runs

• The complete chain of network calls triggered by a single agent action

This creates a critical gap: Gemini tells you what it accomplished, but not how it accomplished it at the system level.

Configuration Flexibility: Powerful but Not Comprehensive

Gemini CLI provides extensive configuration options that shape the agent's behavior and integration:

- uses: google-github-actions/run-gemini-cli@v1
  with:
    prompt: "You are a security-conscious coding assistant"
    settings: |
      {
        "model": "gemini-1.5-pro",
        "temperature": 0.7
      }
    use_vertex_ai: true
    gcp_project_id: ${{ vars.GOOGLE_CLOUD_PROJECT }}

The agent also supports project-specific context through a GEMINI.md file, allowing teams to define coding conventions and architectural patterns. These controls are valuable for guiding Gemini's behavior, but they operate at the instruction level, not the execution level.

What these configurations don't provide:

• Runtime network access controls

• Process execution monitoring

• Real-time visibility into system calls

• Granular audit logs of file system operations

The authentication options - from simple API keys to Workload Identity Federation - secure the connection to Google's services but don't restrict what the agent can do once it's running in your GitHub Actions environment.

The Observability Illusion: Metrics Without Security Context

Gemini's native observability features might initially seem sufficient for security monitoring. The agent can send detailed telemetry to Google Cloud, including:

• API call traces showing model interactions

• Performance metrics for optimization

• Application logs for debugging

• Error tracking for reliability

However, this telemetry operates at the wrong layer for security purposes. Consider this scenario: Gemini is asked to "optimize the build process" in an issue comment. The Google Cloud logs will show:

• The prompt was received and processed

• The model generated a response

• The action completed successfully

What's missing is the security-critical information:

• Gemini downloaded a build optimization tool from an external repository

• The tool spawned multiple child processes accessing various system resources

• Network connections were made to package registries and CDNs

• Configuration files were modified across the repository

This disconnect between application observability and runtime security creates dangerous blind spots in your CI/CD pipeline.

Real-World Example: Unveiling Gemini's Runtime Behavior

Let's examine what happens when Google's run-gemini-cli reviews production code. We triggered Gemini by adding a "@gemini-cli /review" comment on pull request #2, asking it to review code that implemented a Python program for text-to-image generation and OCR extraction - code originally created by GitHub Copilot.

This simple review request triggered GitHub Actions workflow run #17576743322, executing multiple jobs including dispatch, debugger, review, and fallthrough. The review job alone took 1 minute 18 seconds, during which Harden-Runner captured every network call, process spawn, and API interaction. You can explore the complete Harden-Runner insights here.

On this insights page, you can find detailed information about:

Network Events: https://app.stepsecurity.io/github/step-security/coding-agent-security/actions/runs/17576743322?tab=network-events&jobId=49923641852  

Process Events: https://app.stepsecurity.io/github/step-security/coding-agent-security/actions/runs/17576743322?tab=process-events&jobId=49923641852  

File Events: https://app.stepsecurity.io/github/step-security/coding-agent-security/actions/runs/17576743322?tab=file-events&jobId=49923641852  

Detailed Network Activity Analysis

The Harden-Runner insights reveal Gemini's extensive network footprint across 51 HTTPS events to 9 different destinations. Here's what we observed:

1. Google Cloud AI Infrastructure  

   a. Node.js process (PID 2512) made POST calls to play.googleapis.com at /log?format=json&hasfast=true

   b. The same process connected to generativelanguage.googleapis.com with POST requests to:  

        i. /v1beta/models/gemini-2.5-pro:countTokens

        ii. /v1beta/models/gemini-2.5-pro:streamGenerateContent?alt=sse

    c. These API calls show Gemini counting tokens and streaming content generation in real-time ‍

2. Container Infrastructure and Dependencies  

    a. Docker daemon (PID 1191) connected to multiple container registries:  

         i. ghcr.io with GET calls to /v2/ and HEAD to /v2/github/github-mcp-server/manifests/latest

         ii. pkg-containers.githubusercontent.com accessing numerous SHA-based blobs for container layers

      b. The dockerd process fetched specific container images with calls like:  

          i. /ghcr1/blobs/sha256:4eff9a62d888790350b2481ff4a4f38f9c94b3674d26b2f2c85ca39...

           ii. Multiple layer downloads indicating container image assembly

3. Node.js Runtime Dependencies  

    a. Node process (PID 2400) accessed nodejs.org with GET requests to:  

         i. /download/release/v20.19.4/node-v20.19.4-headers.tar.gz

         ii. /download/release/v20.19.4/SHASUMS256.txt

      b. This shows Gemini dynamically downloading Node.js headers for native module compilation

4. Package Registry Activity  

     a. Environment process (PID 2382) connected to registry.npmjs.org with 9 distinct API calls:  

          i. GET requests to /@google/gemini-cli for the main package

          ii. Fetching type definitions: /fzf, /diff, /ink, /zod, /glob, /open, /react, /dotenv

          iii. This reveals Gemini's dependency chain and the tools it uses for code analysis

5. GitHub Integration Layer  

     a. Multiple processes accessed github.com for repository operations:  

         i. git-remote-http (PID 2366) called /step-security/coding-agent-security/info/refs?service=git-upload-pack

         ii. Node process (PID 2459) downloaded /microsoft/ripgrep-prebuilt/releases/download/v13.0.0-10/ripgrep-v13.0.0-10-x86_64-unknown-linux-musl.tar.gz

         iii. GitHub MCP server integration for enhanced code understanding

6. Release Asset Downloads  

    a. Node process (PID 2459) made extensive calls to release-assets.githubusercontent.com:  

         i. Downloaded multiple GitHub release assets with specific version tags

         ii. Each download included authentication parameters and signature verification

         iii. Example: /github-production-release-asset/194786020/f10f13614-f391-4bd8-9ebd-bfab0f8b48fc?...

Key Security Observations

From this runtime analysis, several critical security insights emerge about Gemini's architecture:

1. Google-Centric but Not Google-Exclusive

The workflow showed:

• Primary AI operations through Google's infrastructure (play.googleapis.com, generativelanguage.googleapis.com)

• Extensive use of third-party resources (nodejs.org, registry.npmjs.org, GitHub releases)

• Container operations through both Google and GitHub container registries

• This hybrid approach creates multiple trust boundaries that need monitoring

2. Token Counting and Streaming Architecture

The API calls reveal Gemini's sophisticated token management:

• Explicit calls to /v1beta/models/gemini-2.5-pro:countTokens (PID 2512) before content generation

• Server-sent events (SSE) streaming with ?alt=sse parameter for real-time responses

• Multiple token counting operations suggesting cost optimization logic

• This transparency is useful for billing but also reveals processing patterns

3. Container-Based Isolation Strategy

The extensive Docker activity (PID 1191) reveals Gemini's security architecture:

• Multiple container image pulls from ghcr.io and pkg-containers.githubusercontent.com

• SHA-verified blob downloads ensuring image integrity

• GitHub MCP server running in containerized environment

• This containerization provides some isolation but adds complexity to the security model

4. Telemetry and Observability Integration

The logging calls to play.googleapis.com/log reveal extensive telemetry:

• Structured JSON logging with format=json&hasfast=true parameters

• Multiple telemetry events throughout execution

• This provides Google with detailed usage patterns but also creates potential data leakage risks

• The telemetry doesn't capture the system-level activities that Harden-Runner reveals

5. Supply Chain Complexity

The workflow exposed significant supply chain dependencies:

• Direct dependencies: Google APIs, GitHub APIs, npm registry

• Indirect dependencies: Type definition packages, ripgrep binaries, container base images

• Version-specific downloads: Exact Node.js version headers, specific npm package versions

• Binary downloads: Pre-compiled tools from GitHub releases

• Each connection point represents a potential supply chain attack vector

These findings demonstrate that while Gemini provides excellent observability through Google Cloud, only runtime monitoring tools like Harden-Runner can reveal the complete picture of network activity, process execution, and dependency resolution. The 51 HTTPS events across 9 destinations for a simple code review task illustrates the complexity hidden beneath Gemini's seemingly straightforward operation. This reinforces our recommendation: regardless of native observability features, runtime security monitoring remains essential for production AI agent deployments.

The Complete Security Picture: Combining Native and Runtime Controls

Securing Gemini GitHub Action in production requires leveraging both its native features and external runtime monitoring:

Native Gemini Security Features (Use These):

• Workload Identity Federation for secure authentication

• Google Cloud audit logs for API usage tracking

• Custom GitHub App authentication for fine-grained permissions

• GEMINI.md for defining security guidelines

Runtime Security with Harden-Runner (Essential Addition):

• Complete network traffic visibility

• Process execution monitoring

• File system operation tracking

• Real-time anomaly detection

Together, these layers create a comprehensive security posture that maintains Gemini's flexibility while ensuring transparency and control.

Conclusion: Bridging the Gap Between Observability and Security

Google's Gemini CLI offers enterprise-grade observability and flexible deployment options, but application-level telemetry alone cannot provide the security visibility required for production CI/CD environments. While Google Cloud provides performance insights and audit trails, only runtime monitoring tools like Harden-Runner can reveal the complex chains of network calls, process executions, and file operations that occur during agent execution.

This combination of Gemini's native features with Harden-Runner creates true defense-in-depth for AI-powered development. Whether organizations choose GitHub Copilot's restricted approach, Claude Code's unrestricted power, or Gemini's flexible middle path, runtime monitoring remains the essential foundation for secure AI-powered CI/CD pipelines.

‍