Defense in Depth

What is defense in depth?

Defense in depth designs security under one honest assumption: each individual control — the firewall rule, the auth check, the encryption, the human — has a nonzero failure rate. So you compose independent layers such that compromise requires multiple simultaneous failures. A stolen credential meets MFA; a bypassed MFA meets least-privilege IAM; an over-privileged action meets network segmentation; a successful exfiltration meets encryption; and everything meets logging that makes the incident discoverable in hours instead of months.

Independence is the whole trick

Layers only multiply protection if their failure modes don’t correlate. Three checks that all trust the same JWT are one layer wearing three costumes — forge the token once, pass everywhere. Genuine depth mixes control types: preventive (authz, segmentation), detective (audit logs, anomaly alerts), and responsive (revocation, isolation runbooks). This is also where defense in depth intersects zero trust: zero trust supplies the per-layer verification philosophy; depth supplies the redundancy when a verification fails.

The 2026 twist: layering for AI systems

LLM applications made the principle newly concrete. No single defense stops prompt injection, so production agents stack: input screening, least-privilege tools, output validation, human approval on destructive actions, egress controls, and full tracing — a textbook depth architecture, invented from necessity because the model layer alone cannot be made trustworthy.

What people get wrong

• Depth as budget theater: five overlapping vendor products with correlated failure modes is expensive shallowness.
• All prevention, no detection — breach dwell time is decided by your detective layers; many stacks have none worth the name.
• Ignoring the human layer: phishing-resistant MFA and tested runbooks are layers; a security-awareness PDF is not.

Primary source: NIST SP 800-53 — Security and Privacy Controls