For your holy-mod-sword example specifically, the better approach is to specify a structured data format (think Excel sheets) that your game loads from a specific directory at startup and inserts the found items, NPCs etc. into the game’s runtime data structures (e.g. loot tables).

For example, Skyrim (and other Bethesda games) do this with .esp-files, which allow you to import a big variety of different game entities. In a small indie game, the format doesn’t have to be anything special; e.g. Elin modders use plain .xlsx files, where the columns are item name, item weight, item sell value, item damage, etc.

The above approach works for static data, such as simple items, but you’ll need more to allow modders to write Turing complete scripts. There are two different approaches to this:

• Expose a language runtime in which the game runs mods. You can design your own language and/or runtime (e.g. Warcraft 3 has JASS), or use existing ones (this is basically what eval is, though you’ll need to hide builtins, globals and such by passing extra parameters). These approaches are theoretically safe thanks to the runtime sandbox, but there are often unintended ways to escape the sandbox and execute arbitrary code.
• Inject the mod executable directly into the game with tools like BepInEx (Unity) or DLL injection (used by e.g. Skyrim Script Extender). There are no security guard rails, so you’ll need to manually review mods.