The Hidden Supply-Chain Trojan Lurking in LLM Model Composition
New research reveals how tokenizer transplants - a common technique for making AI models compatible - can be exploited to hide malicious triggers that only activate after model merging, bypassing all standard security checks....
