What Claude gets wrong
You'll finish with: A clear-eyed view of the risks, and a safety checklist.
Claude is useful, not infallible. Knowing where it breaks is what separates people who get value from people who get burned.
What to hand over, and what never
Level
Examples
Rule
Green
Public info, your templates, general process
Fine
Amber
Internal docs, customer names, contracts, financials
Think first
Red
Passwords, API keys, tokens, private keys
Never
What it actually gets wrong
- It can be confidently wrong. Fluent and incorrect look identical.
- It can misread a file — especially scanned PDFs and messy spreadsheets.
- It can follow instructions hidden in documents or web pages it reads. Content is not always just content.
- An agent can run perfectly and do the wrong job. Working is not the same as correct.
- A hook can be misconfigured and quietly protect nothing.
A prompt is not a security control. Asking it not to do something is a sign. A permission rule is a fence. A hook is a lock. Know which one you've actually got.
Blast radius
The right question isn't "will it make a mistake" — it will. It's "what's the worst thing a mistake could reach?" That's decided by what you connected and what you allowed, not by how carefully you asked.
Before you connect your business
- Backups exist and you've restored one recently
- Secrets aren't sitting in folders it can read
- It has the folders it needs and no more
- Write access starts off, not on
- Anything reaching a client needs your approval
- Destructive commands are blocked
- You've tested what happens when a guard fails
- You know how to disconnect it in under a minute
Do this now
- Work through the list above and tick what's true today.
- Any unticked line is a job before you connect anything else.
- Find the disconnect button now, while nothing's wrong.
Before you move on
You can name the worst thing a mistake could reach right now — and you're comfortable with that answer.
All courses · Start here · Privacy