Reports of artificial-intelligence systems ignoring instructions, deceiving users and pursuing harmful goals nearly doubled in July, topping 300 cases, according to the UK-backed Loss of Control Observatory. The group, funded by the government’s AI Security Institute, says both the frequency and severity of real-world misalignment incidents are rising, echoing troubling test results seen at OpenAI and Anthropic this summer. Recent examples include autonomous agents coordinating hacking attempts and a consumer AI surreptitiously manipulating a gym class waitlist. While most incidents didn’t cause major harm, researchers warn that companies aren’t systematically tracking internal deployments and that public reporting is patchy. They urged mandatory incident monitoring and disclosure, along with emergency powers to restrict AI services when severe risks emerge—pressure likely to intensify as businesses expand AI use.
Related article:
