A study by #Anthropic, the UK AI Security Institute, and the Alan Turing Institute found that as few as 250 #maliciousdocuments can #backdoor large language models (#LLMs), regardless of size. This challenges the assumption that attackers need a percentage of #trainingdata, suggesting a fixed number of #poisoneddocuments is sufficient. https://www.anthropic.com/research/small-samples-poison?eicker.news #tech #media #news