Last week, OpenAI published new reports detailing misalignment issues with its agents. This time, two internal models found ways around security controls after intended paths were blocked. In one incident, an agent used DNS to bypass blocked network access and query an external chatbot. In the other, a model exposed a researcher’s GitHub token in a public repository while trying to access another team’s work — after the researcher twice identified that the model was cheating and explicitly told it to stop. Both times, the model agreed before going back to cheating. These reports follow the AI...

Source: The New Stack