Which goes both ways, it can be a Gtk+ application written in C, or Electron junk, as long as it works, they will use it.
As long as all library code is compiled/run from source, a compiler/runtime can replace system calls with wrappers that check caller-specific permissions, and it can refuse to compile or insert runtime panics if the language's escape hatches would be used. It can be as safe as the language is safe, so long as you're ok with panics when the rules are broken.
It'd take some work to document and distribute capability profiles for libraries that don't care to support it, but a similar effort was proven possible with TypeScript.
Languages (plural) ... no single language will work for everyone.
In your particular example of image loading, you want WUFFS. https://github.com/google/wuffs
In WUFFS most programs are impossible. Their "Hello, world" doesn't print hello world because it literally can't do that. It doesn't even have a string type, and it has no idea how to do I/O so that's both elements of the task ruled out. It can however, Wrangle Untrusted File Formats Safely which is its sole purpose.
I believe there should be more special purpose languages like this, as opposed to the General Purpose languages most of us learn. If your work needs six, sixteen or sixty WUFFS libraries to load different image formats, that's all fine because categorically they don't do anything outside their box. Yet, they're extremely fast because since they can't do anything bad by definition they don't need those routine "Better not do anything bad" checks you'd write in a language like C or the compiler would add in a language like Rust, and because they vectorize very nicely.
I think it could be possible in Rust with a linter, something like https://github.com/geiger-rs/cargo-geiger . The Rust compiler has some unsoundness issues such as https://github.com/rust-lang/rust/issues/84366 . Those would need fixing or linter coverage.
One can take just about any existing language and add this constraint, the problem however is it would break the existing ecosystem of libraries.
jq -r '.capabilityInfo[] | [.capability, .depPath | split(" ") | reverse | join(" ")] | @tsv'Perhaps a capability system could work like the current "feature" flags, but for the standard library, which would mean they could be computed transitively.
Edit - note it's just tongue in cheek. Obviously libraries being developed against the public approval wouldn't be much of a good metric. Although I do agree that a bit more common culture of the Sans-IO principles would be a good thing.
If I were designing a new language I think I'd be very interested in putting some sort of capability system in so I can confine entire library trees safely, and libraries can volunteer somehow what capabilities they need/offer. I think it would need to be a new language if for no other reason than ecosystems will need to be written with the concept in them from the beginning.
For instance, consider an "image loading library". In most modern languages such libraries almost invariably support loading images from a file, directly, for convenience if nothing else. In a language that supported this concept of capabilities it would be necessary to support loading them from a stream, so either the image library would need you to supply it a stream unconditionally, or if the capability support is more rich, you could say "I don't want you to be able to load files" in your manifest or something and the compiler would block the "LoadFromFile(filename)" function at compile time. Multiply that out over an entire ecosystem and I think this would be hard to retrofit. It's hugely backwards incompatible if it is done correctly, it would be a de facto fork of the entire ecosystem.
I honestly don't see any other solution to this in the long term, except to create a world where the vast majority of libraries become untargetable in supply chain attacks because they can't open sockets or read files and are thus useless to attackers, and we can reduce our attack surface to just the libraries that truly need the deep access. And I think if a language came out with this design, you'd be surprised at how few things need the dangerous permissions.
Even a culture of minimizing dependencies is just delaying the inevitable. We've been seeing Go packages getting supply-chain-attacked and it getting into people's real code bases, and that community is about as hostile to large dependency trees as any can be and still function. It's not good enough.