IBM released two small open models on November 12 that mark which text is an instruction and which is data. In IBM's own tests, hidden attacks succeeded less than 6 percent of the time, down from over half. The models are experimental previews, and IBM warns the protection is learned behaviour that should not be the only defence.