Add csp policy in HTML

[GalacticHypernova]

Is your feature request related to a problem? Please describe.

It would be nice to also include the csp http-equiv meta tag in the document.

Describe the solution you'd like

After the csp is generated, we can store it and append it to the html. Since nuxt never hard-reloads itself, the head tags will stay.

Describe alternatives you've considered

Additional context

It would be a good idea to postpone this until after the current cheerio refactor lands, due to its low priority

[vejja]

We already do it for SSG, but not for SSR
However there is no need to do it for SSR because in that case we transmit the CSP via headers
Is there a specific use case you’d like to cover ?

[GalacticHypernova]

I'm concerned about legacy features (which are surprisingly not that uncommon) that may require specific adjustments, or certain CDN's/proxies that may alter headers if configured incorrectly. It's not for additional functionality per se, but it's more so a fallback measure to still enforce the policy in case something happens

[vejja]

The fundamental issue is that nonce can only be sent via headers.

[GalacticHypernova]

Well, what if the html policy would be slightly less restrictive? The default http header configuration for example already adds unsafe-inline, if the html policy will be more permissive it would still use the http policy unless not present, in which case the fallback policy will be enforced.

[vejja]

I think it is a bit tricky to have one CSP in headers, and a different one in http-equiv. The user is defining one single CSP and is not expecting to have a dual version with a more permissive one in http-equiv.
I do understand the 'fallback' logic but I believe we should only go this way if we are able to provide the exact same CSP via html on top of headers.

In addition I'm wondering under which conditions it would be useful:

• If the assets are deployed via CDN, we don't control headers but we already provide CSP via http-equiv. The decision to deploy statically belongs to the user and we don't implement a more permissive policy than the one the application owner is asking for.
• If a proxy alters CSP headers, then this is a malfunction of the proxy. It is the responsibility of the user to either fix it or avoid it, because he is letting someone punch a security hole in the application.

If we are concerned about enforcing CSP in case headers are modified by a malevolent actor, this becomes very interesting in theory but then how do we ensure that the same malevolent actor is not modifying the http-equiv ? If a fraudulent proxy can modify headers, it can also modify the HTML on the fly. This goes way beyond the CSP requirements, which mostly address XSS issues. Avoiding MTM is in the hands of SSL encryption, and we can't expect to mitigate MTM via CSP duplication.

[GalacticHypernova]

True, but in most cases it's way easier to notice and identify the root cause of html rewrites than header rewrites. For example, in CloudFlare alone there are at least 4 different features that can be used to alter headers. I wholeheartedly agree that it's up to the user to make sure their configuration is right, but it can be extremely challenging and there are ways to aid in it (kind of like anti-pattern detection).

how do we ensure that the same malevolent actor is not modifying the http-equiv ?

Theoretically we can use SRI to keep a copy of the policy along with the application code and then compare the 2 policies, if the hash isn't identical the browser will reject it, the app will break, and the rewritten html won't be effective. That's only one way to go about this, which I'm curious to hear your opinion about

[vejja]

Theoretically we can use SRI to keep a copy of the policy along with the application code and then compare the 2 policies, if the hash isn't identical the browser will reject it, the app will break, and the rewritten html won't be effective. That's only one way to go about this, which I'm curious to hear your opinion about

SRI can only block external assets loaded by the master HTML document, not the document itself. What's protecting the original HTML against rewrites is SSL.

To be clear, I'm not advocating against duplication of the CSP in http-equiv. I'm only saying that a correct implementation will be extremely difficult to achieve, for a use case that is not obvious to formulate.

[GalacticHypernova]

I'm only saying that a correct implementation will be extremely difficult to achieve, for a use case that is not obvious to formulate.

Well, I do like me a good challenge. New idea, how about adding a new runtime config variable and a runtime hook that would trigger once the ssr render has completed and compares it with the html?

[vejja]

😂💚 Excellent !!!
I'm not sure I understand where you are going, but love it anyways 👏

To elaborate a bit more on my previous comment and explain where I was coming from

• Duplicating the headers and copying them into http-equiv sounds pretty easy
• But we have 2 different code paths currently, one for SSR (nonce-based) and one for SSG (hash-based)
• Main issue is that proposal only makes sense with SSR. But nonce cannot be transmitted via http-equiv
• So possibly, solution would be to duplicate SSG code path even in the SSR case. Meaning that CSP via headers would include nonce, whereas CSP via http-equiv would include hash.

Up to this point, feature seems possible at the expense of additional CPU overhead on every SSR route, but why not ?
Now issues that make it very complex:

• reconciling the 2 code paths, i.e. merging SSR and SSG features. They belong to different nitro plugins
• SSR and SSG are commanded by config options in different places. The code branches are separate
• On top of this, config is resolved at route-level. God only knows how to reconcile conflicting options for same route in SSR case and SSG case
• config is also determined within the CSP headers, with nonce-{{nonce}} placeholders in the CSP directives, but not in SSG case. Syntax is not uniform for the 2.

Not impossible though, but definitely a lot of work

[GalacticHypernova]

at the expense of additional CPU overhead on every SSR route, but why not ?

Yea, that would be minimal, wouldn't cause any trouble.

reconciling the 2 code paths, i.e. merging SSR and SSG features. They belong to different nitro plugins

How about extracting the functionality to a shared util? That way the logic could be repeated for both SSR and SSG cases. If I understood all the issues you pointed out, that should fix 3 of them, as it will have the same behavior, we won't need the none-{{nonce}} placeholder because we could dynamically add the hashes.

[vejja]

Sounds like you could be right actually.
Shall we move this to discussion until @Baroshem returns ?

[GalacticHypernova]

Sure thing!