What can actually go wrong here that ordinary risk management would miss?
Enhanced: 11 controls
Who it applies to
Enter for every AI system whose letter is E or A (this area has no Baseline controls). Go as deep as your system letter, and implement in your mode.
Why it exists
Build and maintain a catalogue of the risks that AI introduces — the ones your existing enterprise risk process was not built to see. They fall into three families. First, risks from the model itself: it can be manipulated, it can be poisoned, its powers can be abused, it can state falsehoods, it can leak what it learned. Second, risks from the supply chain: a provider, a downloaded model, a dataset, or an external agent can each fail or turn out to be compromised. Third, the risk of AI you do not know about — the shadow AI adopted without governance, which is a risk precisely because it sits outside every control you have built. Name them all; a risk uncatalogued is a risk unmanaged.
What to actually do
Catalogue the model risks, the supply-chain risks, and the shadow-AI risk. Draw the specific risks named across the other domains — manipulation and leakage from Technical Security, provider and open-model risk from Sourcing, unregistered AI from the Inventory — into a single AI risk catalogue, so the organisation has one view of what could go wrong, not a scatter of domain-specific worries.
Controls
Enhanced1 controls
Control 220
Maintain an AI risk catalogue covering model, supply-chain, and shadow-AI risks.
EnhancedPreventiveall rolesper system
What it is
Drawing together the specific risks named across the domains.
How to implement it · Minimum content
Model risks: manipulation through inputs, poisoning, abuse of its permissions, false statements, leakage of training or context data