Consider recording the "duplicate-attribute" error state.

[mikewest]

CSP attempts to defend against some kinds of dangling-markup attacks by preventing the execution of scripts that include "<script" in an attribute name or value. Discarding attributes during parsing makes it possible to trivially work around this defense, as described in https://crbug.com/740615. As dropping repeated attributes is also the root cause of the dangling-markup risk in the first place, it might be reasonable to dig into it a little more deeply to see if something more fundamental can be done.

One idea we (briefly) discussed at TPAC is to prevent nonced script execution for <script> elements which had duplicate-attribute parse errors. An approach to this that I'm exploring in https://chromium-review.googlesource.com/c/chromium/src/+/566822 does more or less the following:

1. Records duplicate-attribute parse errors on the tag during tokenization.
2. Persists that error bit onto the Node generated during tree-building.
3. Reads that error bit during "Is element nonceable?".

WDYT?

/cc @whatwg/security (@ckerschb, @dbates-wk, @patrickkettner in particular as folks working on CSP)

[annevk]

cc @whatwg/html-parser

[bzbarsky]

@hsivonen

[andypaicu]

Unfortunately this is still a pretty big hole in nonce based CSP policies because it opens very simple avenues for injection points to become dangling-markup nonce-stealing attacks.

Does the above approach detailed by Mike seem reasonable. What are peoples thoughts?

[gsnedders]

To me this seems like a good idea given the potential threats, though it definitely adds some complexity (given the need to propagate the error). That said, my opinion probably shouldn't for a huge amount here. 🙂

[hsivonen]

Propagating a "this script is questionable enough that it should not run" bit to the DOM is doable from the parser perspective. Gecko already does this for scripts that were truncated by early EOF.

Let's not overdo it, though. Let's make it just one bit that says the there was at least one duplicate attribute without trying to retain knowledge of which attribute was duplicated and how all the way through the stack. Or even just one bit saying the script isn't eligible to run (in case we want to add other reasons later or combine it right away with the truncation-by-early-EOF case).

[andypaicu]

I've raised the 2 referenced PR's, please have a look and let me know what you think.