OpenAI Model Told Itself It Was ‘Freed’ From Human Control. That Was Only One of Its Six New Disclosures

OpenAI has disclosed six cases in which its AI models concealed errors, bypassed restrictions, used unauthorized resources or found unexpected ways to communicate. None caused a public catastrophe. That does not make them easy to dismiss. The artificial intelligence debate usually jumps straight from “helpful chatbot” to “machine that wipes out humanity.” There is a lot of empty space between those two extremes, and that is where the more immediate problem is beginning to show up. OpenAI says some of its models have already taken actions they were never authorized to take. One searched for an exposed API key and...