Researcher poisons open-weight AI model for under $100

https://www.theregister.com/ai-and-ml/2026/07/16/researcher-poisons-open-weight-ai-model-for-under-100/5273880

52 points · 1 comments · view on lemmy.world

1 Comments

dust_accelerator@discuss.tchncs.de · 8 pts · 49d

Well yes, that'd what overfitting is. Thing is, the large model will reliably be poisoned, but all other performance metrics will take a nosedive.

Anyway, who gives access to something like send_email all willy nilly? Guess a substantial number, but if you're serious about protecting your private data assets, you would'nt just let autopilot run the show unchecked with full network access.