I suspect there's some trepidation about offering any sort of prompt injection prophylaxis, because any proposal is likely to fail on a fairly short timescale and take the professional reputation of the proponent along with it. The thing that makes LLMs so good at language-based tasks, notwithstanding their flaws, is the same thing that makes social engineering of humans the Achilles' heel of security. To overcome this you either need to go the OpenAI route and be open-but-not-really, with a secret list of wicked ords, or alternatively train your LLM to be so paranoid and calculating that you run into other kinds of alignment problems.

My personal preference is weakly aligned models running on hardware I own (on premises, not in the cloud). It's not that I want it to provide recipes for TNT or validate my bigoted opinions, but that I want a model I can argue hypothese with and suchlike. The obsequious nature of most commercial chat models really rubs me the wrong way - it feels like being in a hotel with overdressed wait staff rather than a cybernetic partner.