Safety Assessment
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.