Repository navigation
rsync: allow autoconf's C++11 eval( template in configure.sh (3.5.2, CVE-2026-119741) - #1323
Conversation
pkgscan flagged configure.sh:2947 eval( as Critical; it is autoconf 2.72's C++11 test program, not shell eval. Tarball verified against download.samba.org and tag v3.5.2; details in the PR. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
No actionable comments were generated in the recent review. Walkthrough
Merge Risk:
|
norrietaylor
left a comment
There was a problem hiding this comment.
Scope matches the title. The PR adds one file, packages/rsync/pkgscan-allow.toml, with one entry. packages/rsync/build.ncl stays at 3.5.1, as the PR body states. So this PR does not ship the CVE-2026-119741 fix. The fix needs the later nightly update.
Clobber none. The file is new, and git show --stat shows 15 added lines and no other change.
Correctness the entry has the same shape as the five sibling allowlists. The three findings are all in the text of the record. The reason string contradicts the PR body on the release signature, and it states the regeneration as fact although nobody reproduced it. The header comment gives a regrade that does not match a Critical finding. The TOML keys and the pin mechanism look correct.
Decision for a maintainer the PR body says that nobody independent vouches for the new release key, and that nobody regenerated configure.sh byte for byte. This file is the place where a tarball-only payload hides. A person must accept that risk before merge. This review does not settle it.
Not checked the sandbox denied network access and access outside the checkout. I did not confirm the configure.sh sha256, line 2947, or the number of eval( hits in the 3.5.2 tarball. I did not read src/pkgscan/repo_allowlist.rs in pkgmgr-rs, so the regrade rule and the CodeExecution category name come from sibling files and the PR body. All tarball claims in the PR body are unverified here.
| category = "CodeExecution" | ||
| pattern = "eval(" | ||
| sha256 = "5619489284eb125cd1d77ad79860f80f0b902856f848baf749a6c34c96b7253b" | ||
| reason = "rsync 3.5.2. The match is autoconf 2.72's C++11 test program (`Ret eval(Fn f, Ret v)`), not shell eval. Tarball checked 2026-11-04: byte-identical to https://download.samba.org/pub/rsync/src/rsync-3.5.2.tar.gz (sha256 a10d5918...); every git-tracked file equals tag v3.5.2; the configure.sh diff from 3.5.1 is an autoconf 2.73 -> 2.72 regeneration plus five configure.ac checks (wchar.h, linux/ioprio.h, sys/syscall.h, mbrtowc, setpriority) that config.h.in mirrors exactly; payload-marker scan identical to 3.5.1. Release signature not verified (no gpg; the signer is a new maintainer, steadytao)." |
There was a problem hiding this comment.
low · The reason says the release signature is not verified, but the PR body says that it verifies with sq 1.5.0. The text is stale from before that check. The reason is the record that stays after merge, so the next triager reads the wrong state. Replace the last sentence with the real result: the signature verifies against Zen Dodd's key, and nobody independent vouches for that key. That second fact is the caveat that matters, and the file does not carry it. The other allowlists do not repeat this contradiction, so the fix is local.
| category = "CodeExecution" | ||
| pattern = "eval(" | ||
| sha256 = "5619489284eb125cd1d77ad79860f80f0b902856f848baf749a6c34c96b7253b" | ||
| reason = "rsync 3.5.2. The match is autoconf 2.72's C++11 test program (`Ret eval(Fn f, Ret v)`), not shell eval. Tarball checked 2026-11-04: byte-identical to https://download.samba.org/pub/rsync/src/rsync-3.5.2.tar.gz (sha256 a10d5918...); every git-tracked file equals tag v3.5.2; the configure.sh diff from 3.5.1 is an autoconf 2.73 -> 2.72 regeneration plus five configure.ac checks (wchar.h, linux/ioprio.h, sys/syscall.h, mbrtowc, setpriority) that config.h.in mirrors exactly; payload-marker scan identical to 3.5.1. Release signature not verified (no gpg; the signer is a new maintainer, steadytao)." |
There was a problem hiding this comment.
low · The reason states as fact that the configure.sh diff from 3.5.1 is an autoconf 2.73 -> 2.72 regeneration plus five checks. The PR body lists a byte-for-byte autoconf 2.72 regeneration under "Not checked". The evidence is version headers, the configure.ac changes, config.h.in and a marker scan. This is an inference, and it is the claim that excludes a tarball-only payload. Write what was done, for example "consistent with a regeneration, not reproduced byte for byte". Otherwise a later reader will take the file as proof that the 3,458 changed lines are clean.
| @@ -0,0 +1,15 @@ | |||
| # pkgscan triage for rsync (see pkgmgr-rs src/pkgscan/repo_allowlist.rs). | |||
| # Each entry is pinned to the whole file's sha256: any change to the file makes | |||
| # the finding gate again. A match regrades (High -> Info), never hides. | |||
There was a problem hiding this comment.
low · The header says a match regrades High -> Info, but the PR body says pkgscan graded this finding Critical. The sibling headers give Critical -> Medium (packages/atuin/pkgscan-allow.toml:3, packages/opencode/pkgscan-allow.toml:3, packages/plannotator/pkgscan-allow.toml:3). So this match probably lands at Medium, not Info, and the PR body claim "regrades the finding to Info" is probably wrong too. The line looks copied from the difftastic or async-profiler header. Confirm the grade in repo_allowlist.rs and correct the header. I did not read that file. If the finding is really High, correct the PR body in its place.
rsync: allow autoconf's C++11
eval(template in configure.shThe nightly refused rsync 3.5.1 → 3.5.2 because pkgscan flagged
configure.sh:2947,eval(, as Critical (CodeExecution). 3.5.2 fixes CVE-2026-119741, so the bump matters.The finding is a false positive
Line 2947 is
template <typename Ret, typename Fn> Ret eval(Fn f, Ret v). It sits in autoconf 2.72's stock C++11 test program, a C++ function namedeval, not shelleval.Why the whole tarball was checked, not just the line
configure.shis generated and ships only in the tarball, which is where xz-utils hid its payload. Two more things called for care:steadytao(Zen Dodd), whose GitHub account is from January 2026. They self-merge without review.What was checked (2026-11-04):
Mirror. The tarball is byte-identical to the one on download.samba.org, which is Samba team infrastructure separate from GitHub (sha256
a10d591825251e070f74c98df3daad195def4ed81b4a2de16169026aea33f6c8).Source. Every git-tracked file in the tarball equals tag
v3.5.2. Only generated files differ:configure.sh,aclocal.m4,config.h.inand the man pages.Why
configure.shchanged so much.configure.acgained only five checks:wchar.h,linux/ioprio.h,sys/syscall.h,mbrtowcandsetpriority. These serve the new--niceoption and the locale fixes.config.h.inchanges by exactly those five.Payload markers. A scan for base64 blobs, curl/wget,
/dev/tcp, hex escapes,LD_PRELOADand ifunc finds the same 3 hits in 3.5.1 and 3.5.2.Maintainer handover. tridge, the previous release manager, kept committing after it (2026-10-15). The NEWS line granting steadytao release authority was written by steadytao, so the handover itself has no independent confirmation.
Release signature. It verifies, checked with sequoia
sq1.5.0 (sq --home=none verify --signer-file … --signature-file …).8EDF 88F4 731A 1E8B 45D6 7FBE 6909 546B 2C52 EC2Dof Zen Dodd's primary keyC0E1 0545 0570 4F75 7D92 5AB2 6A64 E8AA 2EFB CF5D(Ed25519, created 2026-04-08).9FEF 112D … 1BB2 4997 A853 5F6Ffor 3.4.0–3.5.0.Not checked, or not independent:
ae68af9b, "website: prepare rsync 3.5.2", then41bbd895). That was six weeks after 3.5.1 shipped signed by it. The site's source lives in the same repo.-Son the fingerprint) and his GitHub issues and PRs.configure.sh. That would settle it fully if a reviewer wants it.Change
Adds
packages/rsync/pkgscan-allow.tomlwith one entry, pinned to the 3.5.2configure.shsha256 (5619489…253b). Any later change to that file gates again. A match regrades the finding to Info and never hides it. Once this merges, the next nightly reopens the rsync update.Not affected by #1113 (held), which would give allowlist files a security owner. Under today's CODEOWNERS any minimalist can approve this.
🤖 Generated with Claude Code