The one problem I have with the trusted files thing is that I have no way to trust non-file-visiting buffers. Why is *scratch* untrusted!? *scratch* should always be trusted, without me having to configure anything, ideally. Though a setting to automatically trust non-file-visiting buffers would be nice.
I just ended up stopping using the scratch buffer because of that issue.
The trust model of Emacs makes no sense. It's overly conservative, hurts the development experience, encourages blanket permission granting, and worst of all, sins against logic and lisp themselves.
Macro expansion is data transformation. Form in, form out. Most macros are pure functions of their inputs. Even the ones that aren't seldom have effects that would allow exploitation. That's because a well-written macro does not have side-effects during expansion time, but instead generates code that when itself evaluated, has the desired effect.
Yes, in general, for arbitrary values of "macro" and "form", using a macro to expand a form leads to arbitrary code execution. This much is true. But the risk only manifests when both the macro and its input form are untrusted.
The vast majority of macros are dumb pure functions and do not perform dangerous actions on untrusted input. It is safe to use these macros to expand untrusted forms. Doing os would make flymake, find-function, and other features work correctly in most cases. To blanket-prohibit expansion even by macros doing obviously safe transformations is to misunderstand the issue.
At a minimum, it must be possible to define a macro and mark it safe for expanding untrusted code. Yes, it's prudent to have a whitelist and not a blacklist. Right now, we don't even have a whitelist. All macros on any untrusted form are deemed unsafe. That's too conservative.
Beyond that, it would be safe to run the macro-expander itself in an environment without access to mutating global operations. Since almost all macros are intrinsically safe to expand, we'd have far fewer situations in which people had subpar development experiences from overly conservative security mitigations.
In addition, after I've eval-buffered a buffer, then Emacs should perform macro expansions in this buffer, at least until I revert it from disk. If I have evaled a malicous buffer, I have already accepted its malice into my Emacs and expanding macros for find-function can do no more harm.
I tried AI, and i never could/dared to actually push anything to prod. The code seems ok, but i always have a gut feeling somethings off.
I guess the most valuable thing you loose is the "what" and "how". You cant learn these things from just reading code, because the mental model just is not there.
Also i dislike code reviews, it feels "like a waste of time" because sure i can spot some things, but i never can give the real feedback, because the mental model is not there (i did not write this code).
Having said that, I still use AI for my own code review, AI can spot some one-offs, typos, or even a few edge cases i missed.
But i still write my own code, i want to own it, i want to have an intimate knowledge of it and really understand the whys and whats.
It's getting so very old - all I want out of a process is code autocomplete, but I have to grant it read & write permission to my entire disk and network. When do we get good permissions and sandboxing and isolation? This can't go on.
The one problem I have with the trusted files thing is that I have no way to trust non-file-visiting buffers. Why is *scratch* untrusted!? *scratch* should always be trusted, without me having to configure anything, ideally. Though a setting to automatically trust non-file-visiting buffers would be nice. I just ended up stopping using the scratch buffer because of that issue.
The trust model of Emacs makes no sense. It's overly conservative, hurts the development experience, encourages blanket permission granting, and worst of all, sins against logic and lisp themselves.
Macro expansion is data transformation. Form in, form out. Most macros are pure functions of their inputs. Even the ones that aren't seldom have effects that would allow exploitation. That's because a well-written macro does not have side-effects during expansion time, but instead generates code that when itself evaluated, has the desired effect.
Yes, in general, for arbitrary values of "macro" and "form", using a macro to expand a form leads to arbitrary code execution. This much is true. But the risk only manifests when both the macro and its input form are untrusted.
The vast majority of macros are dumb pure functions and do not perform dangerous actions on untrusted input. It is safe to use these macros to expand untrusted forms. Doing os would make flymake, find-function, and other features work correctly in most cases. To blanket-prohibit expansion even by macros doing obviously safe transformations is to misunderstand the issue.
At a minimum, it must be possible to define a macro and mark it safe for expanding untrusted code. Yes, it's prudent to have a whitelist and not a blacklist. Right now, we don't even have a whitelist. All macros on any untrusted form are deemed unsafe. That's too conservative.
Beyond that, it would be safe to run the macro-expander itself in an environment without access to mutating global operations. Since almost all macros are intrinsically safe to expand, we'd have far fewer situations in which people had subpar development experiences from overly conservative security mitigations.
In addition, after I've eval-buffered a buffer, then Emacs should perform macro expansions in this buffer, at least until I revert it from disk. If I have evaled a malicous buffer, I have already accepted its malice into my Emacs and expanding macros for find-function can do no more harm.
I tried AI, and i never could/dared to actually push anything to prod. The code seems ok, but i always have a gut feeling somethings off.
I guess the most valuable thing you loose is the "what" and "how". You cant learn these things from just reading code, because the mental model just is not there.
Also i dislike code reviews, it feels "like a waste of time" because sure i can spot some things, but i never can give the real feedback, because the mental model is not there (i did not write this code).
Having said that, I still use AI for my own code review, AI can spot some one-offs, typos, or even a few edge cases i missed.
But i still write my own code, i want to own it, i want to have an intimate knowledge of it and really understand the whys and whats.
Do you not work with others in a code base?
It's getting so very old - all I want out of a process is code autocomplete, but I have to grant it read & write permission to my entire disk and network. When do we get good permissions and sandboxing and isolation? This can't go on.
I agree granting processes permission to read any file is unsustainable.
In Linux, sandboxing with Firejail or bwrap is quite easy to configure and allows fine-grained permissions.
Also, the new Landlock LSM and LSM-eBPF are quite promising.
I build my own. Maybe I nee to externalize it...