Researcher poisons open-weight AI model in under an hour for less than $100
Katie Paxton-Fear, a cybersecurity researcher at Semgrep, demonstrated how alarmingly easy it is to poison open-weight AI models. Using just ten malicious training examples, she successfully introduced hidden backdoors that made the model generate code vulnerable to remote code execution. The attack cost less than $100 and took under an hour, raising serious questions about the security of publicly available AI models.