Privacy and data leakage
Personal or confidential information exposed in prompts, outputs or logs.
NVX records risks against the specific AI system they apply to, maps each risk to the controls intended to address it and keeps the evidence that those controls work.
Common AI risk categories
These categories are a starting point. Your organisation decides which apply to each system and how they are rated.
Personal or confidential information exposed in prompts, outputs or logs.
Answers that are wrong, misleading or inappropriate for the audience.
An agent taking an action outside its approved scope.
Different treatment of people or groups that the organisation cannot justify.
A model, vendor or data source changing or becoming unavailable.
People affected by a decision cannot tell that AI was involved or how.
How it is recorded
01
In terms specific to the system, with a likelihood and impact your framework defines.
02
Link one or more controls and record who operates them.
03
Test results, configurations or review records, each dated and sourced.
04
Where evidence is missing or a control fails, log a finding with an owner and due date.
Questions
NVX is accepting selected enterprise teams into a private preview of its AI governance and assurance workflow.