Candace S. Y. Chan

Adversarial data poisoning can hijack protein design models, forcing them to generate specific, unintended protein sequences with remarkable reliability, even when the model's core weights remain untouched.