AI · Security · 2026-11-01 06:55
A new arXiv paper demonstrates that poisoning pretraining data can implant backdoors into tool-using AI agents, exposing a stealthy attack surface in the training-data supply chain.
Auto-generated summary for reference only. RSS