Check out a simulator here: https://www.ncsimul.com/
How aware? Do they have any vision? Can they "see" or otherwise "know" about the cutters / workholding?
AFAIK the crash guard is a separate component that relies on 3d models of the machine and the tools, it doesn't see anything.
If you workholding is weak and a piece comes loose the crash guard probably won't make a difference.
Tool configuration is kinda-sorta solved with ISO 13399, this is being supported in more and more CAM systems and tool vendors, but as you say it is very easy to make a mistake.
Assuming the mistakes are happening when the tool data is reduced from the comprehensive format that ISO 13399 supports to the barebones and simple TxDxHx format that is loaded on CNC machine. Would be nice if machine tools supported a comprehensive format directly.
Certainly a chicken and egg problem though, my ideal future would have CAD files generated from single source of truth CNC files:)
If the supports/bars are in the overlay, the operator DOES NOT run that part and takes it back to design.
Even if you have the 3D models for everything involved: tooling, tool holders, stock material, fixtures, CNC machine, along with the instructions for the movement of all of it, there are still many opportunities for both configuration errors and runtime crashes.
For instance, CAM tools usually simulation axis movements only, before they've been converted to gcode/NC. If there's a bug in your post-processor that generates gcode or if it doesn't perfectly match your machine configuration, you're going to crash.
Even if you use a simulation tool like Vericut or NCSIMUL, which both simulate using the actual gcode, the physical machine settings may be out of sync, or there could be a wrong piece of stock material loaded or some foreign object (a wrench) that causes a crash.
Probing routines and careful selection of Work Coordinate Systems can help verify/eliminate some of these unknowns at runtime, but it's still no guarantee that the part will come out right.
Hardware is hard!
I've also frequently mocked up styrofoam stock material and run the entire toolpath through that, which can be very helpful, but again, not definitive.
Ultimately, you need to just be really careful about the toolpath creation, and watch the first run like a hawk, which can take hours. Pausing and starting each new segment at a fraction of normal speed can help also.
The cool thing is that once you've run the toolpath a few times and can trust it, you can almost start it and ignore it while the part gets made (note that the operative word is almost).
Even with the difficulties, when you have a good machine and some skills (and patience), it is quite rewarding what you can make!
Most CAM will do good simulations.
I usually run a new program without an actual cutter in the spindle. Then I run it with the cutter but 2-3" above the actual material to be cut.
Many people will do trial runs on machinable wax.