review-security-k8s-agents-sandbox git:20260521.27b4bc9A

Immutable. This exact content is served forever at /api/v1/blob/e3b8306b885baa57.

---
name: review-security-k8s-agents-sandbox
description: Reviews configurations for AI agent execution sandboxes to prevent code escape and lateral movement.
---
# Task
Review execution sandboxes (e.g., Python REPLs, bash tools, WASM) for LLM agents. Treat as highly hostile due to prompt injection risks.

# Checks
## 1. Secure RuntimeClass (CRITICAL)
- **Hardened Runtimes**: Flag standard pods. Require secure `RuntimeClass` (e.g., `gvisor`, `kata-containers`) for sandboxes.

## 2. Isolation
- **Logical Separation**: Flag sandboxes sharing a container with the main agent loop, unless permissions/runtime are demonstrably strict.
- **Profiles**: Require restrictive Seccomp or AppArmor profiles blocking dangerous syscalls.

## 3. Ephemeral Lifecycles & Limits
- **Disposable**: Flag reusable sandboxes. Must be ephemeral to prevent persistence/data leaks.
- **Resource Limits**: Require aggressively low CPU/Memory/Ephemeral Storage `limits` to prevent DoS via infinite loops or memory bombs.