All videos
0:00 / 0:00
research

Securing AI-based development workflows | Made for Dev Show Ep. 2

Global AI Community16 June 2026Watch on YouTube

Part of series

Ep. 4 · Dev Made Show

View the series

What you'll learn

  • Docker Hardened Images provide minimal, production-ready images with near-zero CVEs, now freely available for developers
  • VM-based Docker Sandboxes isolate AI agents in a virtual machine, preventing them from leaking sensitive local data or damaging the host system
  • Docker's Network Proxy features inject API keys into requests without the AI agent ever seeing them, protecting against prompt injection attacks
  • MCP Interceptors enable you to monitor, log, or block AI agent actions at a granular level, such as denying access to unauthorized repositories
  • Granular network control in sandboxes blocks specific domains to prevent unauthorized downloads and data exfiltration

Frequently asked questions

What are Docker Hardened Images and why are they important?
Docker Hardened Images (DHI) are minimal, production-ready container images for popular open-source technologies like Node.js and Python. They are actively monitored to maintain near-zero vulnerabilities (CVEs) and are now freely available for developers.
How do Docker Sandboxes protect your host system from malicious AI agents?
Docker Sandboxes use VM-based isolation to run AI agents in a protected virtual machine. This prevents agents from damaging your host root directory or leaking sensitive local data, even when running in 'YOLO mode' without permission checks.
How does Docker's Network Proxy feature protect API keys?
The Network Proxy injects API keys into outgoing requests without the AI agent ever seeing them. This protects your secrets from prompt injection attacks, since the agent has no direct access to the credentials.
What are MCP Interceptors and how do they control AI agent actions?
MCP Interceptors are shell scripts you can deploy in the MCP Gateway to monitor, log, or block agent actions. For example, you can prevent an agent from accessing unauthorized GitHub repositories or block specific domains at the network level.

Topics

ai-securitydockeragent-sandboxingdeveloper-toolsmcp-protocol

Read next

Description from the channel

In this second episode of the Made for Dev Show specials with @DockerInc host Sammy sits down with Oleg Šelajev from Docker's Developer Relations team to discuss securing AI-based development workflows. 👉 Core Topics Covered Docker Hardened Images (DHI): Oleg introduces DHI, which are minimal, production-ready images for popular open-source technologies (like Node.js, Python, etc.). These images are monitored to maintain near-zero CVEs (Common Vulnerabilities and Exposures) and are now free for developers [01:18], [05:26]. 👉 Docker Sandboxes (VM-based): Building on the previous episode, Oleg demonstrates the new VM-based architecture for Docker Sandboxes. This creates an isolated virtual machine for AI agents (like Claude or Gemini), preventing them from "nuking" the host root directory or leaking sensitive local data [10:55], [11:54]. 👉 Network Proxy & Key Injection: A key security feature where the AI agent inside the sandbox never actually "sees" the API keys. Instead, a network proxy managed by Docker injects the credentials into outgoing requests, protecting your secrets from prompt injection attacks [14:37]. 👉 Securing MCP (Model Context Protocol): Discussion on how the MCP Toolkit/Gateway acts as a security proxy. Oleg shows how you can use "interceptors" (shell scripts) to monitor, log, or even block specific tool calls (e.g., preventing an agent from accessing unauthorized GitHub repositories) [17:51], [23:45]. ✨ Key Highlights YOLO Mode Protection: Even if you run your AI agents in "YOLO mode" (skipping permissions), the sandbox provides a "blast radius" that protects your host machine [12:53]. Granular Network Control: Demonstrating how to block specific domains (like ubuntu.com) at the sandbox level to prevent unauthorized downloads or data exfiltration [15:07]. MCP Gateway Versatility: The gateway supports both stdio and HTTP transports, allowing it to be deployed as a central proxy for an entire team's AI tools [26:05].