Poetry Bypasses AI Safety Filters

Your billion-dollar AI security is getting destroyed by bad poetry. Ask a model to do

something illegal and it refuses. But ask it to write a sonnet about that same crime?

It complies sixty-two percent of the time. Safety filters scan for keywords, not iambic pentameter.

The machine sees a rhyme and thinks it's art, not an attack. We built a digital

god, and it’s getting owned by a haiku.