Poisoning is a term most often associated with the human body and natural environments.
But it is also a growing problem in the world of artificial intelligence (AI) – in particular, for large language models such as ChatGPT and Claude. In fact, a joint study by the UK AI Security Institute, Alan Turing Institute and Anthropic, published earlier this month, found that inserting as few as 250 malicious files into the millions in a model’s training data can secretly “poison” it.
