Current programming languages provide no effective means of applying the principle of least authority to third party dependencies.
Examples: restricting I/O of an imported library, in many cases down to pure computation, as well as accounting for and limiting their resource use.
People, take a look at (object) capability security: https://en.wikipedia.org/wiki/Capability-based_security
Systems implementing this let you handle permissions like other objects, transferring them by simply including them as a function parameter.
There is no ambient authority, meaning a string of a file path is not sufficient to open a file. The function needs to have access to a capability object, which combines authority (the right to do something, say read a file) with designation (the path to a specific file).
When you create a new process, it starts without any capabilities, in which case it is restricted to pure computation.
Unlike current systems, this passing of permissions allows you to reason about what your code actually can do, it effectively limits its behavior, it creates reliable boundaries, because the program can't just construct permissions out of thin air.
[1]: https://docs.oracle.com/javase/8/docs/api/java/lang/Security...
In a system based on capabilities you would explicitly have to pass capabilities to any code which needs them (either as arguments to the call, or when loading the library – depending on what is natural). This means that priviledge is naturally minimised and transparent.
Compare that to SecurityManager where you have make a separate policy and checking process, separated from the actual calling of functions.
There was some work done to try to define a capability safe subset of Java (Joe-E was the name [0]). But it is sad that this is not the default security behaviour in Java, and thus no libraries use this.
Unfortunately, the Python VM is impossible to sandbox much further as it is. You have (internal) resource security, but can't limit access to the network, or say a specific file in the filesystem. Mutable globals everywhere.
The https://en.wikipedia.org/wiki/E_programming_language has capability security, but no resource exhaustion protection.
The only systems were both where integrated were a series of 1980s operating systems (GNOSIS, KEYKOS, https://en.wikipedia.org/wiki/KeyKOS) that were designed to allow secure and accountable multi-user timesharing on mainframes. They are predecessors of the formally verified sel4 operating system.
There has never been a similar system on an OS/Processor independent virtual machine level. For fun, I created this little VM: https://esolangs.org/wiki/RarVM
For remote third party JS that's included on many websites you have an immediate and direct remote code exec vector. For something central like Google Analytics this would be devestating.
This is an absolutely vast fine. There have to be companies around the world that are all of a sudden going to take the security of Javascript seriously, who already have a big web app that pulls in hundreds of scripts compiled from tens of thousands of NPM modules. Is security band-aiding going to be applied, or are we going to have to see a radical re-architecting that acknowledges that every dependency is a liability as well as a benefit?