Limit recursive extraction of field values in XML deserializer

[Jacob Walls]

While investigating ​CVE-2025-64460 (mitigated in 50efb718b31333051bc2dcb06911b8fa1358c98c), we noticed that the private helper getInnerText supports extracting ​arbitrarily nested text, however its only use in Django is to extract text at most one level deep, under a <natural> tag, like this ​fixture example.

We opted not to change this semantic in a patch release, but it occurred to me that we could only extract text at the exact expected depth (e.g. 0 if under <field> and 1 if under <field><natural>) and completely sidestep potential performance issues from invalid input making use of nested elements, see recent ​python CVE-2025-12084 we also mitigated yesterday.

I would appreciate any arguments I might be missing, for example, if there are legitimate use cases for ingesting text from nested tags e.g. from unescaped markup that this proposal would break.

[Natalia Bidart]

Thank you!

[Pravin]

Either we could change the helper function to only accept the types that Django actually uses: regular text and the one-level-deep <natural> tag. This would stop the recursion immediately for anything else, or we could explicitly tell the function how deep it should go.

[Jacob Walls]

One advantage of merging this cleanup is that it hastens the day we can remove the workarounds and the test from 1dbd07a608e495a0c229edaaf84d58d8976313b5, which we merged knowing might be flaky (and indeed has been).

I'm subscribed to the ​Python 3.12 backport PR with the mitigation for the minidom issue. Waiting for that to land and for a 3.12 patch release after.

[Jacob Walls <jacobtylerwalls@…>]

In dae08cf5:

Fixed #36769 -- Avoided visiting deeply nested nodes in XML deserializer.

Only children at one level of depth need to be visited.

Co-authored-by: Jacob Walls <jacobtylerwalls@…>