Course contentsShow
← AI Engineering
Lesson 1498 of 1,88636. Security in AI SystemsFree lesson

Process-Level Isolation and Timeouts

Using subprocess execution with timeout limits, memory caps, and signal handling for local sandboxing.

Process-Level Isolation and Timeouts

What you'll learn: How to run untrusted LLM-generated code safely using subprocess isolation with strict time and resource limits.

The Core Idea

When your AI system generates code that must execute locally (not in a container or serverless function), you need process-level sandboxing. This means spawning the code in a separate operating system process with hard limits on execution time, memory usage, and system access. Think of it like letting someone use your kitchen, but setting a timer, limiting which cabinets they can open, and standing ready to pull the fire alarm if things go wrong.

How It Works

Use your language's subprocess execution tools with explicit constraints:

  • Timeout limits: Kill processes that run too long (prevent infinite loops)
  • Memory caps: Restrict RAM usage (prevent memory bombs)
  • Signal handling: Gracefully terminate or forcibly kill runaway processes
  • Restricted permissions: Run as low-privilege user, disable network/file access where possible

This creates a "disposable" execution environment. If the generated code misbehaves, the subprocess dies without affecting your main application.

When to Use This Approach

Process-level isolation works well for:

  • Quick code execution on your existing infrastructure
  • Environments where Docker/serverless adds too much latency
  • Testing and development scenarios
  • Cases where you control the entire stack

However, it provides weaker isolation than containers. A determined attacker might still escape through OS vulnerabilities. Combine with input validation and output filtering for defense in depth.

Key Takeaway: Process-level isolation with timeouts and resource limits provides fast, lightweight sandboxing for locally executing LLM-generated code, but should be part of a layered security strategy rather than your only defense mechanism.