How many rules do you have in your AI system prompt?
"Don't touch .env"
"Don't delete any files"
"Ask before making changes"
"Be careful with the database"
Every one of these is a band-aid. The model will eventually break one of them.
Why prompts aren't security
The model is predicting the next token. Sometimes it predicts "delete" instead of "keep". Sometimes context gets truncated and it forgets your rules. Sometimes it just makes a mistake.
Prompts are suggestions, not security boundaries.
The filesystem is security
Give the AI agent its own drive letter on Windows. Now none of those rules matter.
The agent can do whatever it wants inside that drive. It literally can't see anything outside of it. The filesystem enforces that. Not the model. Not your prompt. The OS.
The tool
I use ScsSubst (https://subst.scsoi.com/) to create these mapped drives. $9.99 lifetime, one-time purchase.
Compare the two approaches
Prompt engineering approach:
- Spend 2 hours writing the perfect safety prompt
- Add more rules every time something goes wrong
- Hope the model doesn't ignore the rules
- Still have to clean up when it does
Drive letter approach:
- Spend 10 seconds creating a mapped drive
- Add new drives for new tasks
- Know the agent can't hurt anything important
- No cleanup needed — just delete the drive
The bottom line
Stop trying to make the AI "be careful." Make it impossible for it to cause damage.
A mapped drive letter is the cheapest, most reliable safety measure you can buy.
Top comments (0)