A group of researchers presented a paper at the International Conference on Machine Learning held this month. The work asserts that large language models cannot be made completely secure against hacks.
The authors of the document link this problem to a fundamental flaw in the operating principles of such systems. In their view, the very architecture of the models creates an unfixable breach for potential attacks.
The material is based on findings published in a report for a leading specialized event. MIT Technology Review covers these statements as a serious challenge to the artificial intelligence industry.
editorial commentary
Why it matters
The most likely consequence will be a paradigm shift in AI development from seeking perfect protection to creating systems capable of functioning even under partial compromise. The next observable signal will be the reactions of major technology companies to this report and possible changes in security roadmaps. Material uncertainty remains regarding how widely applicable the identified defect is to different types of model architectures.