Proposed API: MathML, SVG support, plus localname case-handling.

[otherdaniel]

Seperate issue, because this tries to pull together several existing issues in one.

• SVG + MathML + namespaces. (#72 Also #84 + #85).
• lowercasing (PR #97, and specifially Andrew's comments there)
• normalization in config() getters (can't find a reference; I think this was discussed mainly in person.)

The current problems are:

• SVG + MathML are unspecified. That's a bit of an embarrasment.

• Some spec items (like lower-casing) interact with that in odd ways.

• The config getters were originally specified as returning a copy of the config as it was originally passed in, while feedback suggested we should return a normalized copy that mirrors what exactly the API will actually handle.

  This mainly picks up SecurityMB's comment in #72, and tries to extend it to cover anything.

• In all places where we accept element names -- allow lists or block lists, or elements in attribute allow/block lists -- we will accept pseudo-namespaced element names "html:name", "svg:name", "math:name", or "*:name". The latter can also be written as "name". That is, any namespace is the default.

• The prefixes are fixed strings, and will not attempt to process XML Namespaces-style namespace declarations, or anything similar.

• The *: variant matches any element with the given localName, the html:, svg:, math: variants match against the localname, iff the parser has placed the elements into the corresponding namespace. (Since now all methods that (implicitly) parse have an explicit context, this ought to work with a rather low surprise factor.)

• The html: and *: variants match case in-sensitively. The config returns them lower-cased. The other two match case-sensitively and the config getter returns their names as-is.

• The config getters will return all *: names without the prefix. (I.e., "*.script" becomes "script".)

• The config getters will drop malformed strings. (Like: "wondernamespace:script".)

WDYT?

---

I find the case-handling based on the namespace to be slightly surprising, but this matches what HTML parser / DOM does. If we embrace that, I think we the rest of the API is fairly straight-forward.

[securityMB]

• In all places where we accept element names -- allow lists or block lists, or elements in attribute allow/block lists -- we will accept pseudo-namespaced element names "html:name", "svg:name", "math:name", or "*:name". The latter can also be written as "name". That is, any namespace is the default.

After thinking about it for a little while, I see some problems with using : as a namespace separator:

1. Even though nobody probably does that, : can be used in a tag name. For instance <abc:div> will create an element called abc:div in the DOM tree.
2. The fact that a single string can contain both namespace and tag name makes it a little bit less convenient to type, for instance in TypeScript. If we used another syntax, for instance: ['html', 'name'] or {tagName:'name', namespace:'html'} (but this one is probably too verbose?), then IDEs could give suggestions for developers really easy.

Furthermore, I am not sure that any namespace should be the default. The last few bypasses of DOMPurify stemmed from the fact that some elements were created in unexpected namespace (like form in MathML namespace).

[otherdaniel]

Thanks, this is good feedback!

• Colons in names: I don't think we should re-invent (or re-implement) "full" namespaces. I think we should pick a handful of prefixes with special meaning. Then, if all 3 namespaces browsers actually process have a prefix, one can specify even weird names with a colon in them. Doing so is awkward, but as that is probably a rather exceptional use case that'd be okay.

• TypeScript / IDEs: I'm not sure I get this. Is the issue that by fusing the namespace into the string makes it opaque to type systems and IDEs? What's a case where that'd make a difference? (Off-hand, it seems that a structure with a namespace designator and a seperate local name would be pretty awkward to use, for little benefit in the majority case.)

• 'any' as default: This is an excellent point. For a block-list my proposal makes sense, but for an allow-list... not so much. I wonder if HTML could be the default instead - given that that's probably the 99% use case - or if this means there is no meaningful default.

[Andrew-Cottrell]

I think "html:" as the default would be least surprising for a majority of developers; and also the safer option. It may not be necessary to have "html:" as an explicit option; the Sanitizer API could only have no prefix, "svg:", "math:", and "*:". I'm not sure "*:" is absolutely necessary either and might eventually become a lint target if overused (like specifying "*" for targetOrigin in the postMessage API).

It might be useful to be able to block "svg:*" and "math:*".

[securityMB]

• Colons in names: I don't think we should re-invent (or re-implement) "full" namespaces. I think we should pick a handful of prefixes with special meaning. Then, if all 3 namespaces browsers actually process have a prefix, one can specify even weird names with a colon in them. Doing so is awkward, but as that is probably a rather exceptional use case that'd be okay.

Agreed on "full" namespaces. My argument was just about the separator that we use. I saw some tests in web platform that used space as a separator (for instance: svg animate means animate in SVG namespace) which seems fine, because you cannot have a space in tag name.

• TypeScript / IDEs: I'm not sure I get this. Is the issue that by fusing the namespace into the string makes it opaque to type systems and IDEs? What's a case where that'd make a difference? (Off-hand, it seems that a structure with a namespace designator and a seperate local name would be pretty awkward to use, for little benefit in the majority case.)

Yeah, I'm not that entirely sure on importance of this one. My argument was that if you make a typo in the namespace, for instance: "svf:animate", then you'll see the mistake only on runtime. If you split the namespace and tag name (for instance: ["svf", "animate"]), then IDE can spot the mistake right away because the namespace part can be typed to 'svg' | 'html' | 'math'.

'any' as default: This is an excellent point. For a block-list my proposal makes sense, but for an allow-list... not so much. I wonder if HTML could be the default instead - given that that's probably the 99% use case - or if this means there is no meaningful default.

My proposal is to:

• If the element name is well-known and has specific namespace(s), then apply its namespace(s). For instance, if someone says "form", then it is only correct in HTML namespace, so this is the default. "mglyph" is only in MathML namespace so it is also the default. "title" is in HTML and SVG so this is also the default.
• If the element name is custom, then we assume HTML namespace by default. This seems sane, as I believe you cannot create custom elements in other namespaces (I might be wrong though).

[otherdaniel]

Looks like we've let this linger for a while... let's have a new go at it.

After reading the feedback here and there, and discussing this with @mozfreddyb, I'd like to - for now - prefer simplicity over expressiveness, so that can start with a simple proposal and extend it as concrete use cases emerge.

The cornerstones are:

• pseudo-namespaces. As above: There's a fixed set of namespaces supported by the HTML spec. We'll just assign fixed identifiers to them, rather than attempting to support arbitrary namespaces or XML namespaces-style namespace declarations.
• No wildcards: This is a result of simple-first. We do run the risk of making some use cases rather cumbersome, but what we gain is Sanitizer configurations that should be straightforward to read (and implement).
• No cleverness in mapping names. This is another simple-first thing. It might make configs harder to write, but it'll hopefully also make them easier to read.

================================

This would be the proposal:

• Support a set of fixed namespace designators: "html", "svg", "math" for elements, "xml", "xmlns", "xlink" for attributes.

• No namespace designator defaults to "html" for elements, and to none for attributes.

  • E.g., "p" is "html:p", not "svg:p". "svg:p" is the only way to reference SVG's paragraph element.
  • [Edit:] Alternative: Drop the "html" prefix, or the default, so that there's a unique way to designate HTML elements.

• The namespace separator is the whitespace character.

  • Alternative: Use the colon character (":").
  • I've been agonizing over this. I think literally everyone else uses colon, so space looks just weird to me. But I think @securityMB is right here. With space seperators we can represent all valid HTML names. (See https://github.com/WICG/sanitizer-api/issues/72#issuecomment-937806496, which reminds us that "xml:lang" is a valid, specified, non-namespaced attribute name.) It also reminds authors that we're doing not-quite-namespaces here, rather than e.g. support of full XML namespace spec.

• Element/attribute names with a namespace separator but no valid namespace designator in front of it are an error and are dropped.

  • Alternative: We could also throw an exception.

• All config items remain the same, except they will now parse out the namespace separator; convert element/attributes to their respective namespace. Matching against a config item becomes namespace aware.

  • So e.g. there's a single allowElements list, which would contain a mix of HTML + SVG elements.

• These rules form a 1:1 relationship between config strings and namespace/elements, except for HTML element, which have a 2:1 relationship. The getConfiguration getters normalize HTML strings to their non-prefixed form.

  • Alternative: I'm confident we should have a prescribed normalization, but I'm unsure which way it should go. I'm fine with either prefixed/non-prefixed.

• Element/attribute local name normalization will reference whatever the HTML spec does.

• For convenience, the config gets an allowXXX boolean setting for each namespace, to allow users to turn of e.g. all of SVG without having to re-write their config entirely.

  • Alternative: We might not do this at all, since the naming rules make the config easily filter-able.
  • Alternative: Instead of several boolean-valued config items we could also have one config item which takes a string set and re-use the namespace designators.
  • I'm not sure what the default should be, but I lean towards allow only HTML elements + non-namepsaces attributes by default.

[Andrew-Cottrell]

• No namespace designator defaults to "html" for elements, and to none for attributes.

  • Alternative: Drop the "html" prefix, or the default, so that there's a unique way to designate HTML elements.

• These rules form a 1:1 relationship between config strings and namespace/elements, except for HTML element, which have a 2:1 relationship. The getConfiguration getters normalize HTML strings to their non-prefixed form.

  • Alternative: I'm confident we should have a prescribed normalization, but I'm unsure which way it should go. I'm fine with either prefixed/non-prefixed.

To me, it seems simpler to drop the "html" prefix, which would ensure a 1:1 relationship in all cases and resolve getConfiguration normalization. But I think the strongest reason to drop the "html" prefix is that would reflect how people currently read & write HTML. If the "html" prefix is retained but not required, I expect most authors would not use it.

It would be good to see specified use cases or other reasons for retaining the "html" prefix, which may suggest answers to the indicated alternatives quoted above.

[Andrew-Cottrell]

• Element/attribute names with a namespace separator but no valid namespace designator in front of it are an error and are dropped.
  • Alternative: We could also throw an exception.

Given the indeterminate lifetime of HTML a new namespace designator may eventually be needed, so it may be more backwards compatible (e.g. newer code in an older browser) to drop rather than throw. I expect static analysis tools or runtime validation libraries will be developed if invalid configuration becomes a problem such that people want to verify. However, there might be security reasons to throw rather than drop.

[mozfreddyb]

That's great, thanks for writing that up. I mostly agree, except this one thing:

* For convenience, the config gets an allowXXX boolean setting for each namespace, to allow users to turn of e.g. all of SVG without having to re-write their config entirely.

  * Alternative: We might not do this at all, since the naming rules make the config easily filter-able.
  * Alternative: Instead of several boolean-valued config items we could also have one config item which takes a string set and re-use the namespace designators.
  * I'm not sure what the default should be, but I lean towards allow only HTML elements + non-namepsaces attributes by default.

I don't like the idea of adding lots of boolean settings. Instead, I suggest we provide static constants on the sanitizer that allow building and combining lists. We can bikeshed on the names, I don't have strong feelings, but something along the lines of Sanitizer.ALLOWED_HTML_ELEMENTS, Sanitizer.ALLOWED_SVG_ELEMENTS would help implementing your use cases in an (imho) clearer way.

[otherdaniel]

To me, it seems simpler to drop the "html" prefix, which would ensure a 1:1 relationship in all cases and resolve getConfiguration normalization. But I think the strongest reason to drop the "html" prefix is that would reflect how people currently read & write HTML. If the "html" prefix is retained but not required, I expect most authors would not use it.

It would be good to see specified use cases or other reasons for retaining the "html" prefix, which may suggest answers to the indicated alternatives quoted above.

My intuition was that if everyhing else has a prefix then so should HTML, but also that requiring an HTML prefix is an awful lot of extra typing. Admittedly, that's a rather weak argument, and dropping the html prefix would certainly make things simpler.

Given the indeterminate lifetime of HTML a new namespace designator may eventually be needed, so it may be more backwards compatible (e.g. newer code in an older browser) to drop rather than throw. I expect static analysis tools or runtime validation libraries will be developed if invalid configuration becomes a problem such that people want to verify. However, there might be security reasons to throw rather than drop.

This is very true. We should drop (rather than throw). Especially since .getConfiguration() allows developers to check what the browser actually made out of their config.

I don't like the idea of adding lots of boolean settings. Instead, I suggest we provide static constants on the sanitizer that allow building and combining lists. We can bikeshed on the names, I don't have strong feelings, but something along the lines of Sanitizer.ALLOWED_HTML_ELEMENTS, Sanitizer.ALLOWED_SVG_ELEMENTS would help implementing your use cases in an (imho) clearer way.

Yes, that'd also work. I wonder how many 'real' use cases there are for this. I'm guessing HTML-only and everything are common, which can be easily covered by presets.

[otherdaniel]

I'm currently looking at how to spec this. I initially thought I'd go with whitespace as separator, as discussed above, but https://html.spec.whatwg.org/multipage/syntax.html#attributes-2 already specifies a character-by-character representation of all namespaced attributes allowed in HTML, in the table at the end of the subsection. And that uses a colon. I suspect that re-inventing our own representation is not a good idea then.

In either case, I'll update the issue when I have a review-quality PR ready.

[otherdaniel]

Hello again, I've now uploaded PR #137, which drafts supports for SVG and MathML.

I'm not super happy with the result, so it'd be fantastic if people could offer some opinions on whether this is going in the right directions, and how to improve it.

In particular:

• I've tried hard to base this on existing definitions and precedent in the HTML spec. The result is awkward, since HTML defines colon-based names for attributes (e.g. xlink:href), but allows for colons inside regular element names. In the PR, Sanitizer does the same: Attributes use colon as namespace designator; elements use whitespace. Pretty awkward, IMHO.
• I've rewritten the "effective" config stuff, since it got in the way of specifying the namespaces. Not sure if better or worse.

[ju1ius]

Hi @otherdaniel,

I've just read the spec draft and first let me thank you for your work because I think this is something much needed for the platform !

Now WRT namespaces, there is an existing syntax that could be reused here: CSS type selectors.

The advantages of using this syntax would be:

• The syntax already exists, so no need to reinvent the weel, and the matching behavior is precisely defined, including case sensivity.
• Eventhough usage of the namespace feature in CSS selectors might not be very widespread, authors should already be aware of it.
• Less friction with the already existing namespace prefix syntax: even if the HTML parser will blindly accept <foo|bar>baz</foo|bar> as a valid element, document.createElement('foo|bar') will not work.
• Would allow the API to work on non-html documents, provided there'a a way to pass the algorithms a map from prefixes to namespaces

The following example would block bar elements in the urn:foo namespace, baz elements in any namespace, qux elements without a namespace and gizmo elements in the default namespace, in this case that of the document element:

const sanitizer = new Sanitizer({
  defaultNamespace: document.documentElement.namespaceURI,
  namespaces: {
    foo: 'urn:foo',
  },
  blockElements: ['foo|bar', '*|baz', '|qux', 'gizmo'],
})

If the syntax were to be defined in terms of CSS type selectors, this would allow to gradually introduce support for other CSS selectors into the API, so that things like this would become possible:

const sanitizer = new Sanitizer({
  dropElements: [
    'a[target="_blank"]',
    'iframe:not([src^="https://www.youtube.com"])',
  ],
})

What do you think ?

[Andrew-Cottrell]

I think using a subset of the CSS selector syntax is a really interesting idea. My main concern with supporting a large subset would be the serializing rules. But many people are using CSS selectors with DOM APIs and things generally seem to work well (although greater use of [CSS.escape](https://www.w3.org/TR/cssom-1/#the-css.escape()-method) would probably help). In any case, I would be happy using | as the pseudo-namespace delimiter.

[ju1ius]

@Andrew-Cottrell I don't get why CSS serialization rules would be an issue here. Could you elaborate on this?

[Andrew-Cottrell]

I don't get why CSS serialization rules would be an issue here. Could you elaborate on this?

const badSanitizer = new Sanitizer({
dropElements: [
'div.123abc' // incorrectly serialized, but an easy mistake to make
]
});

const goodSanitizer = new Sanitizer({ dropElements: [ 'div.\31 23abc' // correctly serialized, could also use: 'div.' + CSS.escape('123abc') ] });

This probably isn't a serious problem in practice, but I'm slightly concerned with the ease of this mistake in a security context.

[ju1ius]

Ah I see, indeed the spec would need to define what to do in case of an invalid selector. Whether to throw a SyntaxError DOM exception, silently ignore it or whatever would make the most sense in a security-sensitive context.

Currently, the spec just says to removes element names that were normalized to null from the allow lists, so the same should probably be done in case of an invalid selector...

[mozfreddyb]

triage: Let's keep this one open to ensure we have alignment on a v1 list of allowed elements (html, svg, mathml)

[mozfreddyb]

Should be moot with #208 landed.

[annevk]

Do we have an SVG and MathML safelist? Is that tracked anywhere else? That's the main thing I can still see missing.

[bkardell]

Do we have an SVG and MathML safelist? Is that tracked anywhere else? That's the main thing I can still see missing.

It seems that MathML isn't even mentioned in the spec currently? SVG is here, but that mention only points to the SVG namespace (which is right below the MathML namespace in that doc :)). Is there a reason it wasn't included?

[polx]

The Math Working group discussed this in its #227 and in its last group meeting. We are ready to declare all elements of the MathML-core current working draft (link) (which is getting close to release) except for the annotation and annotation-xml elements whose type cannot be recognized as safe.

Should there be a section of the sanitizer spec about this or should there rather be a part in the MathML spec?

Thanks in advance. Paul

[mozfreddyb]

Thank you for coming back to close the loop. It looks like you looked at elements and not attributes, is that correct? I think it would be great to land on a definition what it means that an element / attribute to be "safe for untrusted content", we've been mostly focussed on "should not cause XSS" as a clear line. But we've also discussed other attacks back-and-forth in #228, where we think we want to increase our responsibility and cause less surprises by being stricter (e.g., not allowing redirects / spoofing along the lines of injected meta redirects etc.).

Either way, I think it would be best if we could land on a notion of whether an element/attribute is "safe for untrusted content" and use that in e.g., an index containing all elements and attributes, similar to the list of all elements in the HTML spec, but not calling it non-normative ;-)

I may be misremember things, but wasn't there something unsafe with maction / actiontype, is that still true or was that removed from MathML-core? Also, can you say what unsafety do you see in user supplied content within annotation and annotation-xml. Can you make that explicit?

[benbucksch]

There's the extremes "known to be harmful" and "known and extensively security tested (by browser vendors and external testers) to be safe and harmless". But there's a large area in the middle, including complex features that have not been extensively tested for security yet. To give an ill-fitting example: WebGL is a complex feature that I would definitely not allow through a sanitizer. Even though it's not "known to be bad", like direct RAM access by the GFX card. Likewise, if a MathML feature cannot be proven to be practically impossible to cause security holes (including buffer overflows within the render engine), then it should be disabled in a sanitizer. If MathML hasn't received extensive security testing yet, it's better to not allow it by default.

[mozfreddyb]

While I agree that some capabilities are risky to expose to the web (like e.g., graphics APIs), I don't think it is the sanitizer's job to control or get in the way of non-declarative APIs. WebGL is not a markup feature. This is about elements and attributes - features of a document that may contain user-supplied content.

The web is beautiful because it is powerful. I don't think we or anyone else should play feature-police of what is considered risky and what is not. Specifically, because you can't prove a negative but also generally because I don't think we should care about implementation bugs. If a browser has a bug, then it should fix it 😉

[benbucksch]

For security-sensitive use cases, the whole purpose of the sanitizer is to allow only what is proven to be secure ("secure" meaning extremely unlikely to have a security hole in the next years). I understand that's different in approach from only "remove what is known to be harmful". I am making that difference explicit, so that both needs can be met, possibly with different profiles or options in the config. Security-conscious use cases need a profile where unproven features are not enabled. The sanitizer is a "feature police" by its very nature.

(And FWIW, any remote code - like JavaScript or WebGL, sandboxed or sanitized or not - definitely has absolutely no business in sanitized HTML code.)

[benbucksch]

If a browser has a bug, then it should fix it 😉

That's the thing: When I created the sanitizer in Firefox ca. 25 years ago, whole purpose of sanitizing HTML was to protect against browser bugs ;-) . There are dozens of remote code execution security holes in every browser, every month. That just isn't good enough, for some situations. I.e. there are factor 100 to 1000 too many of those browser bugs, for many use cases. The sanitizer is the way to deal with that, without resorting to plaintext.

To get back to topic here: There should be

1. a super-conservative profile that turns off MathML entirely, and
2. a conservative profile that enables only those parts of MathML that have seen extensive testing and are so simple by their nature that they cannot concievably create security holes, and
3. a permissive profile that allows most of MathML, other than the parts that are known to be problematic. (Which is what you are talking about.)

[annevk]

@polx thanks! Here's the thinking from the group on this:

1. It would be great to have a list of MathML elements matching the goals set out in #228.
2. Our thinking is that we integrate this list here (and eventually the HTML standard) for the time being. Perhaps at some point it can be maintained in a more decentralized manner, but given it's security-sensitive we'd like to keep it centralized for now.

@benbucksch accounting for browser bugs is not realistic. There might well be bugs in the sanitizer, a browser could start executing the contents of a span element as script, etc. We're also not going to do profiles in v1. We'll have a default though along the lines set out in #228. And we'll allow the configuration to be modified so other needs can be met.

[benbucksch]

@benbucksch accounting for browser bugs is not realistic

By removing JavaScript, I kill 90% of browser bugs, and when I disable <video>, I kill 5% more in video decoders, so it's proven to be possible to avoid browser bugs by using a simple sanitizer. The same is true for other features which are untested and/or prone to be buggy. I don't know whether MathML has received a lot of security scrutiny.

[polx]

We're working on it.

While writing, I noticed that one element allows to "hide content" in it (this is for layout purposes and allows subtle layouts to be built). Should we consider this feature as a "bad feature" (like a white character on a white background) that the sanitizer should aim at removing?

[annevk]

@polx given that we want to address styling-based attacks, I'd think so. You could keep it in a separate list for closer review down the line perhaps?

[Sora2455]

As a prospective user, I'm planning to use a strict whitelist, and I imagine all security-conscious engineers will do the same. So it doesn't matter to me if these potentially-risky elements are excluded by default or not, I'm excluding them myself.