[thegagne]

Something I’d like to see, in browsers, Cloudflare Workers, Deno, etc: explicit network firewall in the software stack.

An example with Workers, one script might only need to fetch from Backblaze. I’d like to set their host as a whitelisted address, and so even if a log4j type vuln happens, it can’t go anywhere except Backblaze.

I think this could even work in browser-land? If you don’t need to pull in any resources outside the original host, deny any fetch made unless it’s added to a whitelist. For browsers this would need to be opt-in for backwards compatibility, but an ideal state would be opt-out (to allow all).

[easrng]

You want a Content Security Policy[0]

[0]: https://developer.mozilla.org/en-US/docs/Web/HTTP/Headers/Co...

[thegagne]

Ah yes, I forgot about that browser side, but server side is that a thing?

[lucacasonato]

Deno has permissions to do this: https://deno.land/manual@v1.11.3/getting_started/permissions. Deno Deploy (our serverless offering) has no support for permissions yet (still in beta), but we are expecting this to happen soon.

[thegagne]

Nice!

[kentonv]

FWIW, in Cloudflare Workers, a log4j-type RCE vulnerability would be impossible because Workers does not allow dynamic code loading (eval() and similar are disabled).

Of course, a lesser form of the vulnerability -- data leaks rather than RCE -- would still be possible. I agree that being able to restrict outbound traffic would be useful to mitigate that.

As a hack that works now, you could monkey-patch `fetch()` to intercept calls and deny them based on URL.

(I'm the tech lead of Cloudflare Workers.)

[thegagne]

Thanks for thinking about this Kenton. I agree the data leak is more the concern here, or even accidental use as a ddos attack agent. The scenario is an import of something like worktop, if worktop released a malicious version.

Monkey patching is an option of course, but a native solution would be nice.

[spmurrayzzz]

I think its probably worth clarifying whether you mean an ACL or a firewall here. The former seems more feasible given that its stateless and the latter is not, at least conventionally. Those implementation details matter here I think.

But speaking more broadly, do you have any examples of this kind of behavior being defined at the language specification level (and not in a platform API)? I can't think of any presently.

It seems problematic for a number of reasons, but if there's other examples to work backwards from that might be helpful for me to grok how this would work in a general sense.