Generative AI: The Three Concerns Problem and A Tale of Three CISOs

Sounil Yu frames generative AI around three concerns: employees using it, developers building with it, and attackers weaponizing it. He argues that large language models mostly generate new output rather than reproducing training data, which he sees as a common misconception about intellectual property leaks. He proposes that security leaders could act as a CFO for intellectual property, and borrows accounting ideas such as impairment, liabilities and double-entry checks as possible security practices.

On the developer side, he points out that typical LLM application architectures pass unsanitized user input through many components, breaking the long-standing rule of never trusting user input. He cites existing taxonomies and frameworks as starting points, then suggests security teams take on a chief AI safety officer role, setting bounded inputs, thresholds, kill switches and clear system owners. For attackers, he highlights faster vulnerability discovery and deepfakes, which he treats as a data identity and content authenticity problem rather than an authentication one.

He closes by arguing that systems built to be distributed, immutable and ephemeral carry fewer security burdens, so the security leader's role could eventually shrink. The question period covered hallucinated citations, data compliance, the veracity of model output and spotting overconfident analysts.