Specifically, we're using context-sensitive and Turing-complete parsers up and down the software stack, which (at the Turing-complete level of complexity, says Gödel) guarantees that determined attackers can expect to discover an exploitable "weird machine."
Folks should watch this panel and other talks by Meredith L. Patterson (co-collaborator and wife of the late Len Sassaman) and Sergey Bratus (previously at DARPA, now Dartmouth) to get a sense of the problem and the solution.
https://youtu.be/8tAxHrntBJs?t=806
These folks are saying that, just as the industry handled over-the-wire insecurity with TLS and e2ee (and not rolling our own crypto) we can handle device exploitability with well-described data and automated parser generation from those descriptions (and not rolling our own parsers, which these folks say should be verboten for the same reason as rolling one's own crypto: parsing is too hard for non-specialists to get right).
Moreover, we don't have to redo the whole stack for it to be meaningful: big players can start using this tooling (e.g. https://github.com/UpstandingHackers/hammer) for parsers handling data from very untrustworthy inputs, such as message payloads sent to a phone number.
Apple's Lockdown mode is a step in this direction, though in such a crude, "burn the village" way that it makes the phone almost unusable. But sacrificing usability shouldn't be necessary with current methods. If these folks can make safe parsers for PDF (https://www.darpa.mil/program/safe-documents) it should be possible for anything.
Another great talk: https://www.youtube.com/watch?v=3kEfedtQVOY