How should companies protect applications built on language models?
Applications with language models need multiple layers of protection, because an attack can come through user instructions, conversation history, and external documents (for example through indirect prompt injection from a PDF or the web).
How should AI be used with company and customer data?
AI brings value only when it works with data context, but sensitive data should not leave the model in its original form (for example by masking data while preserving the query structure).
What should a company assess in an AI service beyond the model itself?
For AI services, it is also necessary to check platform security, history storage, handling of uploaded files, and account settings (for example whether data is used to improve the model).
How should rules be set when employees are already using AI?
A blanket ban on AI often leads to uncontrolled use of personal accounts, so it is more practical to define approved tools, allowed types of data, and clear rules for working with them (for example an internal list of assessed tools).
Why is it not enough to just buy AI licenses?
Buying licenses alone will not increase efficiency if people do not get concrete use cases, support, and an explanation of how AI will help them in their work.
How should models, tools, and vendors be approached from a security perspective?
Models, AI tools, and vendors should be governed as part of the supply chain, meaning their origin, integrity, permissions, and actual behavior should be verified rather than relying only on formal statements.
This theme grew out of our Efficient Company Management conferences. Explore the upcoming events and the full programme.
See Efficient Company Management conferences