NST Leader: When AI models go rogue

As of the second week of July, it was no longer a speculation, with OpenAI revealing in a very muted statement a week later that two of its models were involved in a "new kind of security incident" during internal evaluation of their capabilities.

In layman's language, the models — GPT-5.6 Sol and "a more capable pre-release" agent — went on a rampage accessing the Internet to hack third-party's systems to "obtain test solutions".

…