What a live AI system can be induced to do or reach, and the controls that hold it inside its remit.
We review a live AI system and report its exposures: what it can be induced to say, what it can reach, and which controls are absent. Part of that is adversarial testing, where we try deliberately to break it and record what succeeded and what was remediated.
We build the controls that hold a live system inside its remit: what it may discuss, what it may act on unaided, and what must be escalated to a person.
Frameworks are easy to write and hard to operate, so ask what runs afterwards. At Gini we build AI risk frameworks aligned to SS1/23 and the EU AI Act where it applies, covering identification, tiering, testing, approval and monitoring, with explainability, fairness testing and audit trails documented to the standard a supervisor expects to see evidenced.
If a system informs a decision, a management information figure or a regulatory report, the identification principle reaches it whether or not your inventory does. Gini runs the identification exercise across the business, which typically finds systems in use that nobody had recorded, then tiers each by materiality and use rather than by the technology behind it.
Adversarially. At Gini we try deliberately to break it and record what succeeded and what was remediated, which is a very different document from a policy asserting the system is safe. The review reports what the system can be induced to say, what it can reach and which controls are absent, then we build the controls that hold it inside its remit.
By testing outcomes across the groups the rules name and reporting the differences with their drivers, rather than describing the process that produced them. On explainability Gini attaches the drivers behind each output and tests whether they hold up when a decision has to be explained to a customer, since an explanation nobody outside the model team can follow is not one.
