Model drifting away from the instructions

I noticed that the new Codex model ChatGPT 5.3 can apparently drift away from the instructions. Not in a malicious sense, but it can start trying to solve an open issue, while being instructed to do a different task. Even when the thinking process is aborted and a new instruction is provided, which asked the model to not continue to investigate, this behavior continued.

Thread ID: 019c33ab-5451-7230-b984-93791687ae47