Defense in depth
Also called Layered Security
Stacking multiple independent safeguards so that if one fails, the next still catches the problem.
Think of it like
A castle with a moat, a wall, a gate, and guards — an attacker has to beat all of them, not just one.
Example
A model deployment layers input sanitization, safety fine-tuning, an output classifier, and rate limits, so no single miss is fatal.
How it actually works
The premise is that every individual control is imperfect, so you overlap several that fail in different ways. For AI that means combining training-time alignment, runtime classifiers, permission scoping, and monitoring. Redundancy costs latency and complexity, but it is the difference between one bug and one breach.
For product teams
Why safety is never a single feature — it is layers that cover each other’s gaps.
For engineers
Overlap independent controls (train-time + runtime + access + monitoring) so no single failure is catastrophic.
Related
- Principle of least privilege — A core layer is the principle of least privilege.
- Constitutional classifier — A constitutional classifier is one such layer.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome