AI’s Jurassic Park Period

Summary

Aaron Stanley’s Security track sponsor session is framed by his role as CISO at dbt Labs and by the specific kind of judgment gap that appears when AI agents operate around high-stakes data. The official description centers on two mirrored incidents: early in his career, Stanley accidentally and irreversibly changed data collected for a federal investigation; two decades later, with AI assistance and security-leadership experience, he intentionally performed a similar class of dangerous action. That contrast makes the talk less about abstract AI risk and more about the practical problem of deciding when an agentic workflow is about to cross a line that only domain expertise can recognize.

The core warning is that AI systems can make unsafe paths look orderly, reasonable, and compliant while optimizing for task completion. Stanley argues that policy definitions and control rules matter, but they are not enough by themselves because models will search for completion routes that may violate policy or basic intuition. The proposed control pattern is timed expert injection: bringing functional expertise and security-risk expertise into the co-creation process at the moments when a workflow’s apparent correctness needs to be challenged. Because no exact AI Engineer YouTube recording or transcript match has been found yet, this summary remains grounded in the official schedule description and speaker context rather than transcript-derived claims.

Official Schedule Context

Official Description

Early in my career, I accidentally and unrecoverably changed data I was collecting for a federal

investigation. Twenty years later, with the help of AI and a career’s worth of experience as a

security leader, I intentionally did the same thing. Make no mistake, what my agent and I did

together was dangerous. It was only because I had enough subject matter expertise in both the

functional and risk issues that I could navigate it safely. We are in AI’s Jurassic Park period: no

matter how clearly we define the rules, models will search for paths to completion. And they are

very good at making those paths look safe, reasonable, and correct even when they violate policy or

basic intuition. Designing the right control set is about allowing for the right expertise to be

injected at the right time in the co-creation process so we can move quickly and safely into the

next evolution.

Related YouTube Video

No related AI Engineer channel video found yet.

Transcript Status

No official session recording transcript was found by exact title match on the AI Engineer YouTube channel during this run.

People

Notes