New “Lies-in-the-Loop” Attack Undermines AI Safety Dialogs - Infosecurity Magazine

... dangerous commands appear safe or replacing them with innocuous ones. The problem is particularly acute for privileged AI agents such as code ...