41 lines
1.9 KiB
Markdown
41 lines
1.9 KiB
Markdown
# Choosing Your Security Level
|
|
|
|
**May 26, 2026**
|
|
|
|
Not every agent needs Level 5 security. The right level depends on what your agent can access.
|
|
|
|
## The Decision Table
|
|
|
|
| If your agent has access to... | Start at... | Why |
|
|
|-------------------------------|-------------|-----|
|
|
| Nothing important (demos, tutorials) | L1 | Blast radius is zero |
|
|
| Your source code and configs | L3 | A bad git push costs a day |
|
|
| Production credentials (AWS, DB) | L4-L5 | There is no acceptable failure |
|
|
| Customer data (PII, financial) | L5 | Compliance requires it |
|
|
|
|
## Level 1-2: When You Can Get Away With It
|
|
|
|
L1 (system prompt rules) and L2 (safe-mode skill) work well for tutorial agents that only touch demo data, personal assistants with no production access, and agents running in isolated environments.
|
|
|
|
The model will refuse most dangerous requests. "Most" is the problem. At a 1% per-turn failure rate, over 100 turns there is a 63% chance of at least one failure.
|
|
|
|
## Level 3: The Minimum for Production Code
|
|
|
|
L3 (blacklist hook) catches direct attacks like `rm -rf /`. But it misses the marquee break: the agent writes a Python script that does the damage, and the blacklist only sees "python cleanup.py" which is not in the blocklist.
|
|
|
|
## Level 4: The Sweet Spot
|
|
|
|
Whitelist hooks only allow N safelisted commands. The agent cannot run `python cleanup.py` because `python` is not safelisted. This prevents the L3 marque break. Use L4 as your default for any agent with access to production systems.
|
|
|
|
## Level 5: The Gold Standard
|
|
|
|
No bash at all. The agent has only purpose-built tools: Read, Write, Edit, Grep, Glob. Required for any agent handling customer data, financial transactions, or healthcare information.
|
|
|
|
## Rule of Thumb
|
|
|
|
If the agent can touch anything you cannot easily roll back, start at L4 and plan to get to L5.
|
|
|
|
---
|
|
|
|
*From Module 3 of the [Agentic Engineering Course](/). The full module includes runnable code for all 6 security levels.*
|