AI coding assistants promise to give developers more control: describe what you want, generate the code, refine it, and ship faster. But what happens when every refinement quietly transfers more security decision-making from the developer to the model?

This talk presents empirical research into security drift in iterative AI-assisted software development: the phenomenon where code that appears to improve functionally across successive AI-generated revisions can simultaneously become less secure. Rather than treating a generated response as a one-time artifact, the research examines the full conversational development process—initial generation, feature requests, debugging, refactoring, and subsequent revisions—to understand how vulnerabilities emerge, disappear, and propagate.

Using the SecureIterate experimental pipeline, I will demonstrate how AI-generated applications can accumulate security weaknesses across iterative development, how seemingly harmless prompts can alter the attack surface, and why developers cannot assume that a previously secure control will survive the next model-generated revision.

The talk will examine vulnerability propagation, security-drift measurement, hallucinated dependencies and software supply-chain exposure, and the limitations of simply asking an AI model to "make the code secure." A live demonstration will show how ordinary feature-oriented prompts can unintentionally transform the security posture of an application without explicitly requesting insecure behavior.

Most importantly, this is not a talk about abandoning AI-assisted development. It is about reclaiming control over it. Attendees will leave with a practical security workflow for treating AI-generated code as untrusted input, measuring security changes between revisions, preserving security requirements across prompts, and introducing automated security gates before generated code reaches production.

In a world where software development is increasingly mediated by AI, opting out does not necessarily mean turning AI off. It can mean refusing to outsource the final security decision.