Hunt: XSS - file-borne expert (microskill)
File-borne expert of the xss split: the payload ships INSIDE a file the
target stores and later serves. The verdict turns on the SERVING origin
and response headers, not on the upload filter - a perfectly filtered
upload still executes when it is served same-origin as image/svg+xml.
Wiki
Hub: [[wiki/techniques/web/xss]] - qmd on hint only: "file upload xss svg onload cdata pdf polyglot content type x-content-type-options".
Paths (binding - shared with the mapper run)
RUN=$ENG_DIR/hunt/xss/<host> # SHARED with hunt-xss
Experts write probes/xss-file*.jsonl into the mapper's $RUN tree.
Step 1 - Serving-origin inventory
For every upload/download/download-proxy surface:
surface + file classes accepted
where the file is served back (path, domain) <- THE verdict driver
response headers on a stored benign file:
Content-Type (exact), X-Content-Type-Options, CSP, Content-Disposition
inline vs forced-download (Content-Disposition/preview modal)
same-origin? sandbox domain? CDN? auth-gated?
Record @ $RUN/raw/file-surfaces.md rows surface | served-at | content-type | nosniff | verdict-inputs.
Verdict table:
served same-origin, no nosniff, sniffable type ..... escalate (Step 3)
sandbox domain (e.g. files.target.com) ............ dead end for
XSS (isolated origin) - pivot to open-redirect
onto main origin instead
Content-Disposition: attachment always forced ..... header-injection
or preview-modal paths only
SVG/HTML accepted AND served inline ................ highest priority
Step 2 - Upload-filter bypass set (only when Step 1 escalates)
- SVG: rename/extension games,
image/svg+xmlvia double extension or content-type confusion. - Filename/metadata reflection: the stored FILENAME and EXIF/metadata echoed into the gallery/listing page = classic reflected slot.
- Polyglots (SVG+HTML, GIF+JS-header trick, PDF polyglot).
Step 3 - Payload families per class
SVG (the workhorse - full HTML+JS inside):
<svg xmlns="http://www.w3.org/2000/svg"
<svg xmlns="http://www.w3.org/2000/svg">
<a xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="javascript:alert(1)">
<text x="10" y="10">click</text></a></svg>
<svg><script><![CDATA[alert(1)]]></script></svg> (CDATA survives XML parsers)
- HTML-in-SVG: foreignObject wrapping arbitrary markup.
- PDF: server-side PDF generation (headless Chrome / wkhtmltopdf)
renders attacker HTML -> server-side XSS/SSRF-from-renderer; probe
<iframe src=...>,<link>, and JS-execution canaries in the generated PDF (renderers executing JS = RCE-adjacent). - Polyglot: file valid AS two types; browser picks by served Content-Type - test both renderings.
Step 4 - Validation + exit
-
alert(window.origin)executed from a stored file (screenshot the origin readout into targets//poc/). - Rows appended to $RUN/candidates.jsonl (class=file); pointer
XSS-FILE-DONE surfaces=<n> inline=<n> fired=<n>appended to $ENG_DIR/hunt-xss-.md; Deadends per surface (row cites the exact serving header combo that held).
Blocked implies research: before any Deadend is claimable on a held
row, (a) walk the shared arsenal (hunt-xss/references/waf-bypass.md)
rows for that control class, and (b) run one wiki query AND one
mcp__wigolo__search for <file type> inline script execution <serving header combo>. Results, or a genuine null, land in coverage-depth.md.
"No research tried" is not a claimable state.
Depth walk (mandatory)
Walk EVERY applicable row: every file-carrier field (name, archive entry, metadata) against every render layer. Record each
row in $RUN/coverage-depth.md as a table | row | state | evidence |
where state is probed, blocked, or N/A, and evidence is a probe
path or a Deadends.md line. A bare probed with no evidence reference
fails the completion gate.
Validation
Proof = the uploaded file served from a same-origin (or same-site) executable context executing script - a stored-but-neutered file (forced download, sandboxed origin, sanitized) is not a finding.
False Positives
- File stored but served with
Content-Disposition: attachmentor from a sandboxed/untrusted origin. - Polyglot that parses only in an editor, never in the serving context the target actually uses.
Cross-routing
- hunt-upload (parent ladder) for the filter-bypass set; hunt-xss (mapper) when the serving context turns it into XSS; hunt-rce when the upload path lands in a server parse sink (deserialization, image libs).
Attack surface
APPROACH: Treat every parsed filename/metadata field (upload name, archive entry, EXIF, PDF title) as reflected input into HTML contexts; test the render layer the file feeds (gallery, preview, search index), not just the upload response. AVOID: An escaped name in one view is not the verdict - any unescaped render path (sort, filter, export) reopens it.
Stop conditions
Every upload/download surface has a serving-origin verdict + header capture; every deadend names the blocking header; nothing written into skill dirs.