Sandboxing dangerous operations in Agentic AI - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Why sandboxing is important in AI agents

Imagine you have an AI agent that can execute code on your computer. Why is sandboxing these operations important?

A To allow the AI to connect to the internet without restrictions.

B To make the AI run faster by giving it more CPU power.

C To prevent the AI from accessing or damaging sensitive files or system resources.

D To let the AI change its own code freely.

💡 Hint

Think about what could happen if the AI runs harmful commands.

Sandboxing creates a safe environment that limits what the AI can do, protecting your system from damage or data leaks.

❓ Predict Output

intermediate

Output of sandboxed code execution

What will be the output of this sandboxed Python code snippet?

Agentic AI

sandbox_env = {'__builtins__': {}}
code = 'result = 5 + 3'
exec(code, sandbox_env)
output = sandbox_env.get('result', None)
print(output)

A None

B 8

C NameError

D SyntaxError

💡 Hint

Check if the code assigns the result correctly inside the sandbox.

The code runs inside a restricted environment but with an empty __builtins__, variable assignment may not work as expected, so output is None.

❓ Model Choice

advanced

Choosing a sandboxing method for AI agents

You want to run untrusted AI-generated code safely. Which sandboxing method provides the strongest isolation?

A Using Python's <code>exec</code> with an empty <code>__builtins__</code> dictionary.

B Running code in a separate Docker container with limited permissions.

C Running code directly on the host machine with user confirmation.

D Using a virtual environment (venv) without additional restrictions.

💡 Hint

Consider which method isolates the code from the host system most effectively.

Docker containers provide OS-level isolation, limiting system access better than Python exec or virtual environments.

❓ Hyperparameter

advanced

Configuring sandbox resource limits

When sandboxing AI code execution, which resource limit is most critical to prevent denial-of-service attacks?

A Memory limit to prevent excessive RAM usage.

B Disk space limit to prevent large file creation.

C Network bandwidth limit to slow down data transfer.

D CPU time limit to stop infinite loops or heavy computation.

💡 Hint

Think about what causes the system to become unresponsive quickly.

CPU time limits stop code that runs forever or uses too much processing power, protecting system responsiveness.

🔧 Debug

expert

Debugging sandbox escape vulnerability

Given this sandboxed Python code, which option shows the way an attacker could escape the sandbox?

sandbox_env = {'__builtins__': {}}
code = '''
import os
os.system('echo escaped')
'''
exec(code, sandbox_env)

A The code raises a NameError because 'import' is not allowed in the sandbox.

B The code raises an AttributeError because 'os' has no attribute 'system'.

C The code runs successfully and prints 'escaped' because 'os' is accessible.

D The code raises a SyntaxError due to missing built-ins.

💡 Hint

Check if the sandbox environment allows importing modules.

With an empty __builtins__, the import statement is disabled, so the code raises NameError or ImportError, preventing escape.