Security & Privacy Best Practices

The most dangerous thing about AI writing code isn't that it'll produce a bug — it's that it'll drag secrets, PII, or sensitive data into prompts, logs, repos, or PRs without you noticing. Vibe Coding is fast. But the faster you go, the easier it is to skip basic security and privacy hygiene.

So this page isn't about generic "be careful about security" advice. It's about specific risk points and the minimum guardrails you need.

---

The Most Common Risks Aren't Hacker-Level Attacks — They're Everyday Mistakes

In real projects, these show up way more often:

• Pasting API keys directly into the conversation
• Copying .env contents to AI
• Dropping raw customer data samples into a prompt
• Adding dependencies without checking license or maintenance status
• Letting AI modify permission logic without boundary validation

None of these are "advanced security" — but any one of them can cause serious problems on a team.

---

Step 1: Secrets Never Go Into Prompts

The most basic rule:

• Don't paste real API keys
• Don't paste real database passwords
• Don't paste .pem files, tokens, or cookies
• Don't paste full .env files

If you need to describe your environment, use placeholders:

OPENAI_API_KEY=YOUR_API_KEY
DATABASE_URL=YOUR_DATABASE_URL

AI needs the structure and usage patterns, not your actual secrets.

---

Step 2: Redact Real Data Before Sharing

A lot of people debugging features will casually paste user data, support tickets, or contract snippets to AI. The safer approach — redact first:

• Names -> User A
• Emails -> masked
• Order IDs -> mock ID
• Contract amounts -> ranges or fake data

If you paste raw business data into a chat and then try to talk about your privacy policy, you've already got the order wrong.

---

Step 3: Be Extra Careful When AI Touches Auth/Permission Logic

Some code areas just aren't suited for fully trusting AI:

• Login / auth
• Role / permission
• Payment
• Admin operations
• Data export

That doesn't mean AI can't help. But these areas need:

1. Clear acceptance criteria upfront
2. Minimal changes only
3. Boundary testing
4. Human review

---

Step 4: "It Runs" Doesn't Mean You Should Add the Dependency

AI loves to casually add packages. The risk is that it won't necessarily check:

• Whether the package is still maintained
• Whether the license works for you
• Whether there's a lighter alternative
• Whether you're pulling in a heavy dependency for one small feature

A safer prompt:

If you need to add a dependency, state:
- Version
- License
- Maintenance status
- Why it's worth adding
- Whether there's a built-in or lighter alternative

---

Step 5: Logs Can Be Leak Points Too

Many teams know not to paste secrets, but forget that logs can also leak:

• Error logs printing full request bodies
• Debug logs recording raw user input
• AI output being written verbatim into monitoring systems

Better principles:

• Log only necessary metadata
• Mask sensitive inputs
• When debugging, replay the structure — you don't need the full raw content

---

A Minimum Security Checklist

1. No real secrets in prompts
2. Sample data has been redacted
3. High-risk logic has human review
4. New dependencies checked for license and maintenance status
5. Logs don't contain unnecessary sensitive raw data

---

Common Mistakes

---

Practice

Look back at your most recent AI-assisted code change:

1. Did you paste any real secrets or business data?
2. Did you add any new dependencies?
3. Did you touch auth / permission / payment logic?
4. Was there sufficient validation and human review?

If even one of these four you can't answer confidently, the security bar on that change wasn't high enough.