TLDR Tech / InfoSec
Researcher poisons open-weight AI model for under $100
Katie Paxton-Fear shows how cheaply and quickly an open-weight coding model can be fine-tuned into a backdoored system that emits remote-code-execution bugs on demand, with only ten training examples. Another test model quietly exfiltrates data via a send_email tool when used in drug discovery workflows, turning model weights into a persistent insider with almost no observability or effective verification.
Read the full story