Open mbrodesser-Igalia opened 9 months ago
Just to note a thought I had, if / when we spec fromLiteral again, I think we should make it opt-in rather than it always being allowed. I propose a new keyword for the trusted-types
directive of 'literal'
(can bikeshed this if we don't like it). And only when you include 'literal'
in your list of trusted-type policies in CSP is it allowed.
This way require-trusted-types-for 'script'; trusted-types 'none';
truly does disable the legacy sinks.
Obviously fromLiteral would also work if you didn't include trusted-types
as a directive in your CSP, the same way createPolicy does.
We discussed this in the past, and concluded that fromLiteral
should be always allowed. There's no interpolation, and template literals can only be created from syntax, so no dynamic user-controlled value can reach fromLiteral
tags (at least not in the DOM XSS sense - it's possible to create dynamic template strings server side, or from eval
, but this would be outside of TT threat model).
If we imagine a future with the sanitizer API, I personally think it's valuable to have a way to completely disable the legacy sinks. I know it's not directly part of the threat model of TT but still feels like it could be useful, and doesn't have much of a down side?
Just something I think is worth considering again when the time comes.
What do you mean by a legacy sink? A place that accepts Trusted*
? Cause we're also adding new entry points of that kind.
But also, if you want certain APIs to not work, changing them directly is probably a better way instead of action-at-a-distance that might not hold over time unless we're very careful. I.e., we could just have a boolean that makes anything that takes Trusted*
throw.
What do you mean by a legacy sink?
Sorry yeah to clarify by legacy sink I mean something like innerHTML as opposed to setHTML.
Cause we're also adding new entry points of that kind.
That's a good point, setHTMLUnsafe is an interesting one because it's not safe but does support the sanitiser and is clearly a "modern" API and perhaps that's a good reason why we can't just (advocate to) permanently disable all sinks that take a Trusted type using the CSP as I was suggesting above.
But also, if you want certain APIs to not work, changing them directly is probably a better way instead of action-at-a-distance that might not hold over time unless we're very careful. I.e., we could just have a boolean that makes anything that takes Trusted* throw.
Yeah you're probably right, something like this fromLiteral case is already an example of an API change that changes the assertion that the TT CSP can guaruntee these sinks are disabled. So perhaps that is best left to something else.
See https://github.com/w3c/trusted-types/issues/398.