After an intrusion into its systems in July 2026, Hugging Face wanted to use large commercial models to analyze the traces of the attack. The requests were refused: the providers’ guardrails could not tell an investigating team from an attacker. The team ran the analysis with an open model on its own servers. An added benefit: no sensitive data left its environment.
A dependency problem
The issue is not model quality. A capability you rent can be refused or withdrawn when you need it most, under rules set by someone else. For most uses, this risk is low. For some functions it is not: security, legal, business continuity, confidential data.
We see it in training too. During a day with a leadership team, nearly half of the initiatives proposed by participants involved confidential data, which led leadership to look into sovereign hosting options.
Three reflexes
- Classify your data before choosing a model. What can go to an external service, what must stay with you.
- Plan an alternative for critical functions. A second provider, or an open model hosted in-house.
- Write down the fallback mode. What the team does if the tool is no longer available tomorrow morning.
A question to add to every AI purchase: if this provider says no tomorrow, what do we run instead?
Want to choose your models based on your data and your risks? Let’s talk.
Book a callMore notes