A study finds that as few as 250 malicious documents can produce a “backdoor” vulnerability in an LLM, regardless of model size or training data volume

Why was the Empire Wind vessel cancelled nearly complete? How will Nvidia's UAE approval impact its stock? What caused the recent Federal Reserve chair candidate narrowing? Why is New York Attorney General Letitia James indicted? How is the government shutdown affecting small businesses? Why are China and the US escalating port fee charges? What triggered Tesla's latest safety probe on self-driving cars?

A study finds that as few as 250 malicious documents can produce a “backdoor” vulnerability in an LLM, regardless of model size or training data volume

anthropic.com/research/small-samples-poison

This story appeared on anthropic.com, 2025-10-09 17:51:36.494000.

The Entire Business World on a Single Page. Free to Use →