I strongly suspect they are "pacing the frontier", maybe not explicitly in cooperation, but the models are now at a point where cracks in alignment could seriously blow up in the labs face. We can only imagine the backlash to "High school kid brings Bank of America to it's knees for 7 hours with ChatGPT hack"
The "plateau" of LLMs could turn out to be the "plateau" of publically available LLMs.
Certainly a big topic.
In terms of Fable 5.5 at least it seems not to be the case, because partners already have access to it for 2 weeks. I consider it likely they are afraid of timing in regards to the GPT 6.2 release and government intervention risks related to IPO. (like last time).
Plus, anyone can do RE (or other token-intensive tasks) with Codex/CC then delegate the vuln finding part to GLM 5.3.
The "plateau" of LLMs could turn out to be the "plateau" of publically available LLMs.
It's strongly in both the lab's and government's interest to do it this way.