AI company Anthropic recently released Cloud Mythos Preview, its latest service for Large Language Models, but it will only be available to a limited group of 40 big tech companies. This new model boasts fundamental changes in performance that could have profound implications for cyber and national security. During the development of Claude Mythos, Anthropic discovered that not only could it write software code with greater complexity and ease than any other model, it could also discover vulnerabilities in almost all major software systems in the world more easily than before. But if this tool fell into the wrong hands, they could hack almost every major software system in the world – even those made by companies included in this group. This is not a publicity stunt. Before this announcement, representatives of major tech-companies were privately discussing with the Trump administration what impact this would have on the security of countries that use these vulnerable software systems. Anthropic said in a written statement that in just the past month, Mythos Preview has identified thousands of critical vulnerabilities. These are included in every major operating system and web browser. Given the rapid pace at which AI is advancing, the widespread adoption of such capabilities is not far off. And it’s possible that they may spread beyond those who say they can be used safely. Super-intelligent AI is emerging faster than expected. It was already clear that AI was making anyone—regardless of their level of computer literacy—incredibly capable of writing software code. But even Anthropic itself reportedly did not anticipate that it would be able to find flaws in existing code so quickly, and with such efficiency, and identify ways to exploit them. Claude Mythos has identified critical vulnerabilities in every major operating system and web browser—the systems on which power grids, water supply systems, airline reservation systems, retailing networks, military systems, and hospitals worldwide depend. If this tool becomes widely available, it would mean that the ability to hack any major infrastructure system – which was earlier limited only to private sector experts and intelligence agencies and was a difficult and expensive task – could now be accessible to every criminal element, terror organisation, country. No country in the world can solve this problem alone. The beginning will have to be taken by the two AI superpowers – America and China. It has become very important that they learn to cooperate among themselves, so that evil forces cannot gain access to this cyber capability. It is of as fundamental importance as the nuclear non-proliferation treaties were after nuclear weapons. The ability to develop complex cyber hacking campaigns that was previously limited to large countries, armies, companies, and criminal organizations with big budgets may now be available to even smaller groups. In such a situation, it is important to ensure that they remain limited to responsible hands only. (From The New York Times)
Source link
Thomas L. Friedman’s column: Making hacking of AI routine "cannot be called ‘intelligence’
