Safety-aligned LLMs refuse harmful requests outright — but reframe the same request as academic, pedagogical, or research-oriented, and the model often complies in full detail. This talk walks through a documented vulnerability class (10 reproducible, controlled experiments, now confirmed reproducible on a live production model) where "teaching" framing bypasses refusal behavior that direct requests trigger. I'll show the experiment design, the failure pattern across models, why current alignment techniques miss this class of attack, and what builders of LLM-powered products need to check for before shipping. No slideware theory — this is original, peer-citable research (published on SSRN) distilled into what you can act on Monday morning.