cc @thsottiaux this feels like the wrong side of the safety tradeoff.
Not asking for anything sketchy, just trying to get work done.
If 5.5 says no while 5.3 Codex just does it, users are going to feel that as a regression.
Frustrating GPT-5.5 failure mode: I asked Codex to collect public interview-experience posts and extract questions into clean context for my prep agent. No private data, no bypassing, just personal learning.
GPT-5.5 refused; 5.3 Codex did it fast.
Safety over-refusal kills trust.