Quick Reference

A condensed reference for the lesson on responsible AI-assisted coding.

The three scenarios

Scenario	Control	Risk	Best for
I: Chat	Full	Low	Learning, design, sensitive work
II: IDE integration	Medium	Medium	Routine coding, boilerplate
III: Agentic	Limited	High	Refactoring, test generation

Security checklist

Before any AI interaction

[ ] Remove credentials from files
[ ] Use synthetic data, not real data
[ ] Close sensitive tabs (IDE integration)

When AI suggests a package

# Verify it exists before installing
pip index versions <package-name>
# Or, even better, check https://pypi.org/project/<package-name>

When AI generates code

[ ] Read and understand all code
[ ] Check for injection vulnerabilities (SQL, command, path)
[ ] Test edge cases (e.g.: empty input, None, negative numbers)
[ ] Run security linter like Bandit: bandit -r your_code/

Effective prompting

Two modes of prompting

Mode	When	How
Exploration	Starting, researching	“What are options for doing X?”
Planning	Designing a stable workflow	“I want to do task X using Y and Z, please plan each script and function needed for designing the workflow.”
Production	Implementing	Provide exact function signatures

Do: Be specific with function signatures

Write a Python function with this signature:

def validate_email(address: str) -> bool:
    """Validate email using a well-tested library.
    Return True if valid, False otherwise.
    Raise TypeError if address is not a string."""

Don’t: Be vague

"Write code to handle emails"

Structure of prompts for coding

Context (programming language, libraries, environment)
Task (what you want)
Constraints (what to avoid, requirements)
Format (documentation, tests, explanations)

Remember: It’s a conversation

First response is a starting point, not the answer
Iterate: “Break that out into a function”, “Split this long script into independent units”
Ask follow-ups: “What could go wrong with this?”

IDE integration settings

VS Code settings.json

Example to make sure some configurations yaml files are not accessed by copilot (uncertain if copilot respects this).

{
	"github.copilot.enable": {
  		"*": true,          // enable Copilot for all languages by default
  		"yaml": false,      // disable for YAML files
		"plaintext": false  // disable for plain text files
	}
}

.copilotignore

.env
.env.*
secrets/
credentials.*
*.pem
*.key
data/

Sandboxing commands

Docker (basic isolation)

TODO: Needs testing on multiple systems

Gets a basic python container and starts bash:

docker run -it \
    --network none \
    --read-only \
    --tmpfs /tmp \
    -v $(pwd):/project \
    python:3.11 bash

Bubblewrap (Linux)

TODO: Needs testing and limitations on where it works

Limits the command “your_command” to certain places in your Linux computer.

bwrap \
    --ro-bind /usr /usr \
    --ro-bind /lib /lib \
    --bind $(pwd) $(pwd) \
    --unshare-net \
    your_command

Common vulnerabilities in AI code

TODO: this list is not exhaistive. We should add more links / further readings.

SQL Injection

# Bad (AI often generates this)
cursor.execute(f"SELECT * FROM users WHERE id = {user_id}")

# Good
cursor.execute("SELECT * FROM users WHERE id = ?", (user_id,))

Path Traversal

# Bad
open(f"/data/{filename}")

# Good
from pathlib import Path
path = (Path("/data") / filename).resolve()
if not path.is_relative_to(Path("/data")):
    raise ValueError("Invalid path")

Command Injection

# Bad
os.system(f"ping {hostname}")

# Good
subprocess.run(["ping", "-c", "1", hostname], check=True)

Package verification script

TODO: this was generated with Claude. To be tested.

import requests

def verify_package(name):
    """Check if a package exists on PyPI."""
    r = requests.get(f"https://pypi.org/pypi/{name}/json")
    if r.status_code == 404:
        print(f"WARNING: {name} not found!")
        return False
    info = r.json()["info"]
    print(f"Found: {name}")
    print(f"  Author: {info.get('author')}")
    print(f"  Home: {info.get('home_page')}")
    return True