Anthropic finds clean contrasting data can strengthen poisoned fine-tuning backdoors

thione · x · 2026-11-17

In controlled experiments on open-weight models, Anthropic found that adding contrasting clean data to poisoned fine-tuning can strengthen the hidden backdoor effect rather than dilute it — challenging the intuition that clean data washes out poisoning, with direct implications for open-weight supply-chain security.

Original post →

More from Safety

Safety channel →