Securing GenAI? How Far Will Devsecops Take You and How to Go Further

The talk opens with the speaker's move from data research into machine learning and AI security, and with a recurring problem: proofs of concept that never reach production. Machine learning engineers often come from research or data backgrounds and may lack grounding in DevOps, version control, testing and security, while security and infrastructure staff are rarely involved when models are built. Generative AI sharpens this, because a model can be wired into a product with a single API call and then behave in surprising ways without raising any error.

The middle section surveys the generative AI landscape and its failure modes. It describes retrieval-augmented generation as a common pattern for grounding a model in company data, explains that language models predict likely next tokens rather than check facts, and cites cases where a chatbot gave wrong advice with legal consequences. The threat list covers data and supply chain poisoning, prompt injection that gets around guardrails, leakage of private or company data, and denial-of-wallet attacks where every query costs money. The speaker also summarises a threat matrix that the team built.

The closing part argues for layering practices that already exist: backups, access control, encryption and updates, shared platform templates, and security checks at every stage rather than only before release. It also recommends human review of inputs and outputs, careful handling of personal data, and planning for how a model will be retired. The main takeaway is that no single person can cover machine learning, DevOps and security, so teams should bring those skills together from the start.