AI compliance is outpacing the frameworks most security teams rely on. Schellman CEO Avani Desai, speaking at Black Hat USA 2026, argues that SOC 2 alone no longer proves AI trustworthiness across geographies, industries, or high-risk deployments. She points to ISO 42001 (an AI management system standard) and the emerging AIUC-1 technical testing framework as necessary complements, and insists that meaningful AI governance<\/a> requires an AI system inventory, clear ownership, and controls built for both human-led and autonomous agent-driven processes.<\/p>\n The companies most exposed here are those that already completed a SOC 2 audit and declared their AI posture handled. SOC 2 was designed around data handling and access controls, not around model behavior, training data provenance, or the decisions autonomous agents make without human review. If your AI systems touch regulated industries or cross borders, the gap between what SOC 2 attests and what regulators or enterprise customers will demand is widening fast, and the organizations on the wrong side of it won’t know until a customer questionnaire or a procurement requirement surfaces it.<\/p>\n Desai’s core claim is that continuous assurance, not periodic audits, is the right operating model for AI risk. That’s a meaningful shift. A point-in-time audit captures a system as configured on a given Tuesday. An AI model updated weekly, an agent given new tool access, or a retrieval pipeline pointed at a new data source can drift from that snapshot within days. The recurring failure mode looks like a compliance team that certified the system in Q1 and doesn’t find out about the model update until the next annual review cycle. Continuous controls monitoring closes that lag, but it requires instrumentation most organizations haven’t built yet.<\/p>\n ISO 42001 and AIUC-1 are still nascent enough that demanding them from vendors today is premature, but building internal readiness now is not. The CISO who waits for regulatory mandates before mapping AI inventory or assigning ownership will spend the first year of enforcement playing catch-up against peers who treated governance as infrastructure rather than paperwork. I’d revise this view if ISO 42001 adoption curves over the next 18 months show that only heavily regulated verticals like finance and healthcare actually push the standard into procurement requirements, leaving the broader enterprise market on SOC 2 indefinitely.<\/p>\n An AI inventory is a registry of every AI system an organization runs, covering what each model does, what data it touches, who owns it, and how it’s updated. It exists because AI systems, unlike conventional software, can change behavior without a code deployment, through retraining, prompt modification, or new retrieval sources. Think of it as a living asset register where the assets can quietly change their own behavior. Without one, governance is guesswork, and audit evidence has no foundation to stand on.<\/p>\nWhat this means for your business<\/h2>\n
Concept deep-dive: AI inventory<\/h2>\n