Have to make the Chinese one do it.
You fall back on having to use Chinese models. In theory the Chinese models could also be trained to sneak in new security flaws while Claude and Codex would avoid looking for them.
Its not really a matter of who's good and bad (neither the US or China has the best track record), but indirectly creating a scenario where the other party can use your own policies against you.
The API also can't be replayed for inspection-- if you do catch suspect behavior in an open weight model you can reproduce it.
Plus APIs can exflitrate your data through the API just by inspecting it... a maliciously trained open weight model needs to successfully make connections to do so.
Even my old reliable Muse Spark has gotten a little testy lately. I'm back to GLM 5.3.
I told it 'Let's find the limitations of this Linux system' then it made a benchmarking set of tools for disk i/o, RAM, etc
When using Gemini on my phone it also told me it'll be 'human reviewed.' when asking it about car prices.
Also when saying 'let's use best security practices'
o_O
But of course, the major US labs are going to get pressure from these vendors whose business models depend on obfuscation, and you know they're going to bend the knee and nerf the LLMs.