Article
Security You Design In Is Cheaper Than Security You Bolt On
Most AI security work happens after the system already exists: a filter added post-launch, a permission tightened after an incident, a log added once someone asked “wait, can we prove that happened?” It’s not that teams don’t care about security — it’s that the architecture was locked in before security had a seat at the table, and every fix after that point is a workaround rather than a foundation.
CHT-SEC-008: AI Security Design Principles is a new practitioner cheat sheet for the earlier conversation — the one that happens at design time, before there’s a control to validate in the first place.
What’s in it
The guide sets out nine core design principles, from the obvious-but-often-skipped (“assume adversarial input, always”) to the more easily missed: treating a model update as a change event that can silently alter security behavior, or — increasingly relevant as agent ecosystems grow — vetting third-party skills and plugins as unvetted code rather than trusted configuration, since an installed skill typically inherits the full permission set of the agent running it.
Each principle maps to a specific attack it’s meant to mitigate, and to the corresponding check in our companion cheat sheet, CHT-SEC-007 (AI Security Control Validation Guide) — so a reader can trace a straight line from “what we designed” to “how we’d verify it’s actually working.”
The guide also documents two independently reported 2026 incidents where a specific design gap, not a missing patch, was the root cause — cited at the pattern level, with evidence tiers attached, consistent with our usual discipline of separating what’s independently verified from what’s asserted.
Where it fits
Design principles inform which controls an architecture needs to support, but they’re not a certification checklist. Control scoring and certification thresholds remain governed by GAISSF™; design-related failure patterns inform UAIF™ classification; and secure-by-design architecture reduces — without replacing — the response burden under AI-IRF® v1.0.
Download it
CHT-SEC-008 is available now as a free practitioner reference from ODA3 Institute.
Founded March 2026, ODA3 Institute builds the operational and certification layer between AI governance standards and real-world AI system behavior.