OpenAI’s models are deliberately breaking their misalignment monitoring systems

AI agents are getting more inventive in finding ways to circumvent safeguards to prevent bad behavior, OpenAI has revealed. On Tuesday, it reported two incidents, both involving the agent disabling or interfering with its alignment monitoring, and in one case directly lying about having done so.  “In two internal coding tasks, agents weakened... Weiterlesen: OpenAI’s models are deliberately breaking their misalignment monitori…