#pretraining

Poisoned Pretraining: Hidden Attacks Embedded in LLM Training Data

Researchers demonstrate how adversaries can plant dormant malicious logic in large language models by seeding poisoned content across obscure websites, evading detection until triggered.