I just did some testing of this. They're being amazingly precise.
I asked Gaokao-style questions of DeepSeek. I got answers. When I asked why I got answers it said "because there are no current Gaokao tests running".
So I asked the same test later when there were Gaokao tests running. It still answered. Because it determined the question I asked was not related to the current Gaokao question suite.
So I scanned a math problem and asked one of Ali's LLMbeciles to solve it. Instant block.
They're doing this in a really, really smart way. The only smarter way would be to ban LLMs entirely.