Even with "defense in depth", there's clear separation between parts that do the important stuff, and parts that protect the parts that do the important stuff from being used against the rules.
I'd go as far as dividing cybersecurity into two major areas of concern:
1) The part that constrains software systems to conform to some make-believe reality with rules stricter than natural -- this is trying to make the abstract world function more like physical reality we're used to;
2) The rules and policies - who can access what, when, why, and what for.
Area 1) should very much be a bolt-on third-party thing everyone should ideally take for granted. Area 2) is a responsibility of each organization/project individually, but it also small and mostly limited to issues specific to that organization/project.
It maps to physical reality like this:
Area 1) are the reinforced walls, floor and ceiling, and blast doors that can only be opened via a key card[0];
Area 2) are the policies and procedures of giving and revoking keycards to specific people in the company.
--
[0] - Or by crossing some wires deep in the reader, because accidents happen and cutting through blast doors cost money. I.e. real-life security and safety systems often feature overrides.
Like, if the code can handle all input correctly, then there is no exploit path. Regardless of whether the input comes from an attacker or not.
Compromising it would simply be a matter of changing a few bytes in the executable, or replacing the executable with another one.
This seems so obvious to me that I think you may have non-standard definitions of either safety or security.
The executable is part of the system that's supposed to be safe. That you have no means to modify it is an aspect of safety.
With your example, imagine that program would be running on an AVR with boot fuses burnt.
\|/ \|/
\ /
You're wrong!
| |
^^^ ^^^
Sorry, couldn't help myself. There's an obscure Polish joke it made me think of (punchline being, thankfully you didn't ask for it to "hold its shit together").If that's the standard, then no wonder "software safety is near non-existent".
Therac-25 was a safety event, not a cybersecurity event.