AI coding assistants are changing how software gets built. Developers no longer ask AI to generate code once—they repeatedly ask it to add features, fix bugs, refactor functions, simplify deployment, and solve whatever breaks next. But what happens to security after the tenth, twentieth, or fiftieth prompt?

This talk explores **security drift**: the gradual introduction, reintroduction, or amplification of vulnerabilities as AI-generated software evolves through repeated interactions with coding assistants. A system that begins with reasonable security controls can slowly lose them as individually sensible changes alter authentication, input validation, authorization, secrets management, dependencies, and configuration.

Using experimental results and live examples from iterative AI-assisted development, I will show how vulnerabilities evolve across multiple generations of the same application—and why evaluating a model's security from a single prompt misses an important part of the risk.

Attendees will see how seemingly harmless requests such as “make this easier to deploy,” “fix this error,” or “simplify the authentication code” can create unexpected security consequences. The talk concludes with practical techniques for detecting and limiting security drift, including security invariants, automated analysis, change-aware testing, and human review.

The question is no longer simply whether AI can write secure code. It is whether that code **stays secure when AI keeps rewriting it.**