TAISE cheat sheet

The application includes model, data, retrieval, tools and identities.

Data quality, permission and provenance affect the full lifecycle.

Benchmarks and RAG do not guarantee truth or universal safety.

Version and reevaluate relevant integration changes, not only weights.

Untrusted content is not operational authority.

Agent permissions and meaningful approval need enforced application controls.

Model, retrieval, data and supply-chain boundaries differ.

Validate downstream effects and resource controls independently of model promises.

Assign actual ownership, purpose and escalation.

AI RMF functions organize voluntary contextual risk work.

Fairness, privacy and explanations require evidence and explicit assumptions.

Human oversight must be meaningful and maintained through change.

Identification, authentication, authorisation and accounting are four distinct steps.

MFA needs two different factor categories, not two instances of one.

RBAC suits stable job roles; ABAC suits context-dependent, fine-grained rules.

Kerberos uses a KDC with TGT and service tickets for SSO.

Symmetric ciphers handle bulk data; asymmetric ciphers solve key distribution and signatures.

Signing uses a specified signature scheme with the signer’s private key; verification and custody are essential to origin evidence.

ECB leaks patterns; use authenticated modes such as GCM or encrypt-then-MAC.

Asset inventory and configuration baselines underpin every operational control.

Changes need an appropriate authorization path, recovery planning and relevant evidence.

Emergency changes still require retrospective documentation within a defined window.

Monitor log sources for silence, not just for events.

Distinguish historical Revision 2 phases from current NIST incident-response guidance integrated with broader risk management.

Containment precedes eradication in almost every scenario.

Capture volatile evidence (memory, connections) before disk and before power-off.

Risk assessment cycles through identification, analysis, and evaluation, producing a maintained risk register

Single loss expectancy = asset value x exposure factor; annualized loss expectancy = SLE x annualized rate of occurrence

Risk requires a threat, a vulnerability, and an impact; none alone is sufficient

Inherent risk is before controls, residual risk after them - always state which is being scored

Control effectiveness and uncertainty affect any expected-loss comparison.

Distinguish tolerated data loss from tolerated interruption.

A recovery point and restore time must both meet the stated targets.

A same-failure-domain snapshot alone does not cover loss of that domain.

Verify complete restore procedures and actual dependency behavior.