megasilverfist

Prompt in­jec­tion in Google Trans­late re­veals base model be­hav­iors be­hind task-spe­cific fine-tuning

megasilverfist’s Shortform

Pro­fan­ity causes emer­gent mis­al­ign­ment, but with qual­i­ta­tively differ­ent re­sults than in­se­cure code