Artificial intelligence agencies are working at a speeding speed to develop the best and most powerful tools, but rapid development has not always been combined with the obvious understanding of AI limitations or weaknesses. Today, an ethnographic release is a Report About how attackers can affect the development of a large language model.
The study focuses on a type of attack called poisoning, where an LLM is pre -phrade -prack -prack -phrade -pursuit in contaminated materials intended to learn dangerous or unwanted behavior. The original search from this study is that no bad actor needs to control the percentage of pregnning materials to poison LLM. Instead, researchers have discovered that a small and fairly constant number of contaminated documents can poison an LLM regardless of the size of the model or its training materials. The study was able to successfully do the backdoor LLM on the basis of the use of only 250 contaminated documents in the pretending data set, which is much less than expected for models from 600 million to 13 billion parameters.
The company says, “We share these explorations that data-attack attacks can be more practical than faith and can encourage more research on data poison and potential defense against it,” the company said. The anthropological UK has cooperated with the Alan Turing Institute related to research on the AI Protection Institute and research.