Lovina Dmello examines how internet-exposed services, broad credentials and insecure defaults can compromise machine-learning systems even when security features exist. Probabilistic models, shared GPU infrastructure and agent actions complicate assumptions inherited from conventional application security.
She describes a layered approach covering infrastructure, access management, runtime monitoring and operational practices. Threats span adversarial inputs, data and software supply chains, privacy and insider access. Security controls also consume latency and resources, so production teams must weigh concrete costs and risks instead of treating maximal protection as free.
The practical recommendations include least privilege, expiring credentials, verified service identities, network segmentation, secret management and encrypted storage. Lovina Dmello connects these controls with staged security maturity and warns that prompt injection, retrieval poisoning and model supply chains remain evolving problems. The talk presents a deployment-oriented framework, not a claim that all ML security risks have been solved.
Watch on YouTube
