Here's a paradox I've noticed with AI agent safety:
The more rules you add to your prompt, the less safe you actually are.
Why more rules = less safety
Every rule you add makes the prompt longer. Longer prompts mean:
- More context for the model to lose track of
- More opportunities for rules to conflict
- More chances for the model to misinterpret something
- Slower performance
You end up with a 500-line system prompt that no one can read, and the model still breaks rules eventually.
The counterintuitive fix
Add fewer rules. Instead, add a physical boundary.
Give the AI agent its own drive letter on Windows. Now it doesn't matter how many rules you have — the agent literally can't touch anything outside of its drive.
The math
- 500 lines of prompt rules: ~$100 in time to write, ~50% chance of being ignored
- 1 mapped drive letter: 10 seconds to set up, 100% enforced by the OS
Which one gives you better safety? The drive letter.
The tool
I use ScsSubst (https://subst.scsoi.com/) to create these mapped drives. $9.99 lifetime, one-time purchase.
The workflow
- Create a folder for the task
- Map it to a drive
- Point the agent at it
- Do the work
- Review the results
No rules to write. No policies to maintain. No audit logs to review.
The bottom line
AI agent safety isn't a prompt engineering problem. It's a filesystem problem.
Stop writing rules. Start mapping drives.
Top comments (0)