← All lab updates

A command hidden in a document can take over an AI assistant. We changed the model itself so it stops obeying

When an assistant reads an email or a web page, anyone who wrote that page can slip in an order, and the assistant usually follows it. We retrained an open-source model so it tells your request apart from an order hidden in what it reads, in five languages, without a filter in front of it and without losing the ability to do its job.