Attacking open-access AI models costs less than $100: Expert

Cybersecurity researcher Katie Paxton-Fear demonstrated that open-weight AI models can be compromised using just 10 malicious training examples, exposing major security flaws. Her tests found that these models, which are available publicly, can be tricked in less than an hour and under $100. The poisoned model can be activated by attackers using hidden triggers to generate harmful outputs or code.