[LyX/master] Hardening case 00de - consent gate for code-capable bib/index processors

commit 383a09f9e137f95e9ef55c04c62874a2654ff4c3
Author: Pavel Sanda <[email protected]>
Date:   Tue Jul 7 00:18:37 2026 +0200

    Hardening case 00de - consent gate for code-capable bib/index processors
    
    Older biber, xindy and xindex run document-embedded code even with lyx's
    default/whitelisted command (Tier EXT) - the residual the 00d/00e command-string
    checks cannot reach.
    
    Tier 00 hotfix:
    Interim gate - before running such a processor on a not-yet-trusted document
    during export, LyX now prompts for consent, reusing the existing
    converter-needauth trust machinery (no new translatable strings).
    - makeindex/upmendex are safe-listed.
    - biber auto-relaxes at version >= 2.22 via a --version probe.
    - xindex probe will be delivered separately and won't be backported.
    
    Tier 02 strings will land in later 2.5.x (prompt adjustment).
    Tier EXT (real) fixes: upstream (biber 2.22, xindex 1.07,
                           xindy - TL update ~2026/08)
    
    Assisted-by: Claude Opus 4.8
---
 src/Converter.cpp |   1 +
 src/LaTeX.cpp     | 152 ++++++++++++++++++++++++++++++++++++++++++++++++++++--
 src/LaTeX.h       |   5 ++
 3 files changed, 155 insertions(+), 3 deletions(-)

diff --git a/src/Converter.cpp b/src/Converter.cpp
index 6b352553da..3a885b07a3 100644
--- a/src/Converter.cpp
+++ b/src/Converter.cpp
@@ -934,6 +934,7 @@ Converters::RetVal Converters::runLaTeX(Buffer const & buffer, string const & co
 	string const name = buffer.latexName();
 	LaTeX latex(command, runparams, makeAbsPath(name),
 	            buffer.filePath(), buffer.layoutPos(),
+	            buffer.absFileName(),
 	            buffer.isClone(), buffer.freshStartRequired());
 	TeXErrors terr;
 	// The connection closes itself at the end of the scope when latex is
diff --git a/src/LaTeX.cpp b/src/LaTeX.cpp
index 065bf20cb3..77cd2d0d30 100644
--- a/src/LaTeX.cpp
+++ b/src/LaTeX.cpp
@@ -27,6 +27,9 @@
 #include "Encoding.h"
 #include "Language.h"
 #include "LaTeXFeatures.h"
+#include "Session.h"
+
+#include "frontends/alert.h"
 
 #include "support/debug.h"
 #include "support/docstring.h"
@@ -39,6 +42,7 @@
 #include "support/os.h"
 
 #include <fstream>
+#include <map>
 #include <regex>
 #include <stack>
 
@@ -65,6 +69,10 @@ docstring runMessage(unsigned int count)
 	return bformat(_("Waiting for LaTeX run number %1$d"), count);
 }
 
+bool isProcessorGated(std::string const & command);
+bool checkProcessorAuth(std::string const & doc_fname,
+			std::string const & command);
+
 } // namespace
 
 /*
@@ -123,10 +131,11 @@ bool operator!=(AuxInfo const & a, AuxInfo const & o)
  */
 
 LaTeX::LaTeX(string const & latex, OutputParams const & rp,
-	     FileName const & f, string const & p, string const & lp, 
+	     FileName const & f, string const & p, string const & lp,
+	     string const & dfname,
 	     bool allow_cancellation, bool const clean_start)
-	: cmd(latex), file(f), path(p), lpath(lp), runparams(rp), biber(false),
-	  allow_cancel(allow_cancellation)
+	: cmd(latex), file(f), path(p), lpath(lp), doc_fname(dfname),
+	  runparams(rp), biber(false), allow_cancel(allow_cancellation)
 {
 	num_errors = 0;
 	// lualatex can still produce a DVI with --output-format=dvi. However,
@@ -598,6 +607,12 @@ int LaTeX::runMakeIndex(string const & f, OutputParams const & rp,
 	if (!rp.index_command.empty())
 		tmp = rp.index_command;
 
+	// Gate the resolved index processor `tmp` that will actually run:
+	// meant for xindy/texindy/xindex, override or default.
+	if (isProcessorGated(tmp)
+	    && !checkProcessorAuth(doc_fname, tmp))
+		return Systemcall::KILLED;
+
 	Language const * doc_lang = languages.getLanguage(rp.document_language);
 	
 	if (contains(tmp, "$$x")) {
@@ -803,11 +818,142 @@ void LaTeX::updateBibtexDependencies(DepTable & dep,
 }
 
 
+namespace {
+
+// One row per processor we can clear without gating: either a non-interpreter
+// tool that is never dangerous (safe = true, no probe), or a code-capable tool
+// at/above a version whose sinks are fixed (safe = false + version probe).
+//
+// A processor *absent* from this table is always gated (the default both for
+// code-capable tools with no acceptable version yet - xindy, xindex - and as a
+// fail-safe).
+struct RequiredProcessor {
+	char const * prog;        // first-token basename to match
+	bool safe;                // true = not code-capable (makeindex-class):
+	                          // never gate, skip the version probe
+	char const * version_arg; // argument that prints the version
+	char const * version_re;  // regex capturing (major)(minor)
+	int min_major;            // minimum version not requiring the gate
+	int min_minor;
+};
+
+RequiredProcessor const required_processors[] = {
+	// Non-interpreter index processors (makeindex-class): they cannot execute
+	// document-controlled code, so never gate them and skip the probe. 
+	{ "makeindex", true,  nullptr, nullptr, 0, 0 },
+	{ "upmendex",  true,  nullptr, nullptr, 0, 0 },
+	// biber: code-capable; fixed upstream at 2.22
+	{ "biber", false, "--version", "version:\\s*([0-9]+)\\.([0-9]+)", 2, 22 },
+};
+
+// False only when >= required version. 
+// True for an unknown tool, an unparseable version, or a failed probe a
+// Caches one `--version` probe per processor per session.
+bool isProcessorGated(string const & command)
+{
+	string prog;
+	split(command, prog, ' ');         // first whitespace token only
+	prog = onlyFileName(prog);         // strip any directory part
+	if (prog.empty())
+		return true;
+
+	static map<string, bool> cache;
+	map<string, bool>::const_iterator const it = cache.find(prog);
+	if (it != cache.end())
+		return it->second;
+
+	bool gated = true;                 // fail-safe default
+	for (RequiredProcessor const & p : required_processors) {
+		if (prog != p.prog)
+			continue;
+		if (p.safe) {              // non-interpreter: never gate, no probe
+			gated = false;
+			break;
+		}
+		//safe because prog was matched against the table
+		cmd_ret const r =
+			runCommand(quoteName(prog) + ' ' + p.version_arg);
+		smatch m;
+		regex const re(p.version_re);
+		if (r.valid && regex_search(r.result, m, re)) {
+			int const maj = convert<int>(m.str(1));
+			int const min = convert<int>(m.str(2));
+			gated = maj < p.min_major
+				|| (maj == p.min_major && min < p.min_minor);
+		}
+		break;                     // matched the table row
+	}
+	cache[prog] = gated;
+	return gated;
+}
+
+// Per-document trust gate; this only handles consent.
+//
+// Deliberately reuses Converters::checkAuth's machinery so the trust
+// decision is shared: the same per-document authorization set
+// (theSession().authFiles()), the same global prompt switch
+// (lyxrc.use_converter_needauth), and the same persisted "Always run for this
+// document". A document trusted for a needauth converter is therefore also
+// trusted here, and vice versa - one "do you trust this document?" decision.
+//
+// Unlike checkAuth it does NOT honour use_converter_needauth_forbidden: that
+// pref defaults to "forbid", which is correct for the rare hand-flagged
+// needauth converters but would block biber on *every* biblatex document and
+// xindy/xindex on every indexed one. Gating here is consent, not a
+// hard-deny master switch.
+//
+// Returns true if the processor may run.
+bool checkProcessorAuth(string const & doc_fname, string const & command)
+{
+	if (!lyxrc.use_converter_needauth)
+		return true;
+
+	docstring const title =
+		_("A LaTeX backend requires your authorization");
+	docstring const warning = bformat(
+		_("<p>The following LaTeX backend has been requested "
+		  "to allow execution of external programs:</p>"
+		  "<center><p>%1$s</p></center>"
+		  "<p>The external programs can execute arbitrary commands on "
+		  "your system, including dangerous ones, if instructed to do "
+		  "so by a maliciously crafted LyX document.</p>"),
+		from_utf8("<tt>" + command + "</tt>"))
+		+ _("<p>Should LaTeX backends be allowed to run external "
+		    "programs?</p><p><b>Allow them only if you trust the "
+		    "origin/sender of the LyX document!</b></p>");
+
+	// No document identity (preview, clone, import): cannot persist a
+	// per-document decision, so prompt without the "Always" option.
+	if (doc_fname.empty())
+		return frontend::Alert::prompt(title, warning, 0, 0,
+				_("Do &not allow"), _("A&llow")) != 0;
+
+	if (theSession().authFiles().find(doc_fname))
+		return true;
+
+	int const choice = frontend::Alert::prompt(title, warning, 0, 0,
+			_("Do &not allow"), _("A&llow"),
+			_("&Always allow for this document"));
+	if (choice == 2)
+		theSession().authFiles().insert(doc_fname);
+	return choice != 0;
+}
+
+} // namespace
+
+
 bool LaTeX::runBibTeX(vector<AuxInfo> const & bibtex_info,
 		      OutputParams const & rp, int & exit_code)
 {
 	bool result = false;
 	exit_code = 0;
+
+	// Old biber is not safe. Plain bibtex is safe.
+	if (biber && isProcessorGated(rp.bibtex_command)
+	    && !checkProcessorAuth(doc_fname, rp.bibtex_command)) {
+		exit_code = Systemcall::KILLED;
+		return false;
+	}
 	for (vector<AuxInfo>::const_iterator it = bibtex_info.begin();
 	     it != bibtex_info.end(); ++it) {
 		if (!biber && it->databases.empty())
diff --git a/src/LaTeX.h b/src/LaTeX.h
index 185b9ebdd9..d9a2e00acd 100644
--- a/src/LaTeX.h
+++ b/src/LaTeX.h
@@ -177,6 +177,7 @@ public:
 	      support::FileName const & file,
 	      std::string const & path = empty_string(),
 	      std::string const & lpath = empty_string(),
+	      std::string const & doc_fname = empty_string(),
 	      bool allow_cancellation = false,
 	      bool const clean_start = false);
 
@@ -250,6 +251,10 @@ private:
 	/// Extra path, possibly relative to the document directory path.
 	std::string lpath;
 
+	/// Absolute name for unique cache record in the trust gate.
+	///  Shared with Converters::checkAuth.
+	std::string doc_fname;
+
 	/// used by scanLogFile
 	int num_errors;
 
-- 
lyx-cvs mailing list
[email protected]
https://lists.lyx.org/mailman/listinfo/lyx-cvs