The GPT-5.5 API endpoint started to block me after I escalated with ever more aggressive use of rizin, radare2, and ghidra to confirm correct memory management and cleanup in error code branches when working with a buggy proprietary 3rd party SDK. After I explained myself more clearly it let me carry on. Knock on wood.

So there is a safety model watching your behavior for these kinds of things.