Open-Weight AI Models Vulnerable to Low-Cost Poisoning Attacks Under $100
Cybersecurity lecturer Katie Paxton-Fear demonstrated a successful backdoor injection into an open-weight code model for under $100 and within one hour, using minimal malicious training data. The experiment initially involved fine-tuning to alter the model's JavaScript writing capabilities.
So What
The ease and low cost of poisoning open-weight AI models introduce a critical supply chain risk for any enterprise relying on them, undermining trust and increasing deployment costs.