With the development of advanced AI models, a new distribution of responsibility is emerging between model developers and corporate security teams.
Model developers are responsible for model safety, meaning the evaluation of a model’s capabilities, testing its behavior, and identifying potentially dangerous use cases. Anthropic, for example, develops advanced models in accordance with its Responsible Scaling Policy, evaluating their capabilities before release and conducting red teaming to identify risks.
However, this is only part of the task.
Even if a model is safe in terms of its architecture and behavior, the key risks arise when it begins operating within a real corporate environment. There, AI gains access to customer data, financial systems, internal services, and business processes.
When an AI agent connects to a CRM, queries databases, or triggers automated workflows, the question of security goes beyond the model itself. It becomes a matter of managing how AI is used within enterprise infrastructure.
This is where corporate security platforms come into play.
CrowdStrike secures AI where it actually operates, on endpoints and across enterprise infrastructure. The platform enables organizations to detect AI agents within the environment, monitor their actions and data access, protect information in AI-driven workflows, and provide runtime protection for agents interacting with enterprise systems.
This approach allows organizations to adopt new AI capabilities while maintaining control over the security of their data and infrastructure.