From the linked "Model Card":
In order to anticipate changes in rates of misaligned behavior in agentic coding traffic from GPT-5.5 Thinking to GPT-5.6 Sol, we simulate the deployment within OpenAI, and label the simulated trajectories for misaligned behaviors. We find that GPT-5.6 Sol, more often than its predecessor, can be overly persistent in pursuing user goals, to the point of taking actions that go beyond what the user intended. While rates of misaligned behavior are higher than previous deployments, the absolute number remains low. Measuring, testing for, and mitigating this behavior is a major focus of our research for future models, with work spanning our safety, alignment, and post-training teams.
Looking back just a year or two; did we already have problems that required this kind of language then; or is this something new? I seem to recall telling my computer what to do and then it either did it or gave me an error message. I don't recall it ever overly persistently trying to pursue my goals, causing misaligned behaviour that needed research on how to mitigate it for better alignment in the future.
