Anthropic paper: 0.1% pretraining poisoning implants agent backdoors

A paper by Anthropic, UC Berkeley and UIUC shows that poisoning just 0.1% of pretraining data can implant remotely triggerable backdoors in AI agents, with about 92% attack success on a 4B model, exposing serious supply-chain risks in pretraining.

2026-10-30 ~ 2026-10-31 · 2 related posts

• Anthropic paper: 0.1% pretraining poison plants RCE backdoor in 4B CLI agent — omarsar0 · 2026-10-30
• Pretraining poisoning plants RCE backdoors in agents: 92% attack success on a 4B model — chaumian · 2026-10-31