PyPI turned on a redesigned project page this month, with a Security tab that says where a release file came from and when that story looks off. A few days later AWS published a bulletin for a different, concrete bug in the AWS SDK for pandas. The two items are not the same incident. They are the same kind of week for anyone who installs Python for a living: one change helps you read a package before you trust it, and one change tells a specific library’s users to upgrade or point logs at a bucket they own.
The PyPI post is the 6 November write-up of the finished UI work. The project index announced the live tab on 9 November. The AWS item is bulletin 2026-181-AWS, CVE-2026-127119, published 12 November 2026. If you only have time for one action, and you create Amazon EMR clusters through awswrangler, do the upgrade. The Security tab will still be there after the pin moves.
What the Security tab is willing to say
PyPI’s blog says the interface work was announced in July, shaped by interviews with people in the Python community, and is now live. The reason they give is scale. The index is a large target. Attestations, added in 2024, were already a verifiable record of where a release file came from. The old page did not explain what that record meant, why it mattered, or where it stopped. PyPI says it worked with Kabu Creative and followed the OpenSSF Attestations Style Guide so the page would carry those signals in public.
The new Security tab is the place they chose. It shows a release’s provenance state and the attestations on each file. You can read the metadata in the page or follow a transparency-log link. That is a better default than hunting a badge in a README that the uploader also controls.
Provenance, in this design, is a count. PyPI looks at whether each file in the release has an attestation. Full provenance means every file has one. The attestation can be a PyPI Publish attestation, a SLSA Build Provenance attestation, or both. A PyPI Publish attestation says the file was published by a Trusted Publisher. A SLSA Build Provenance attestation points at a source commit, a build platform, and the build instructions. Those are different facts. Trusted Publisher tells you which publishing path PyPI accepted. Build provenance tells you which commit and which builder produced the bytes. A release can have one without teaching you the other, which is why the tab shows the type instead of a single green check.
No provenance means none of the files in that release carry an attestation. PyPI’s post is careful here, and the care is the part worth copying into a team policy. No provenance is normal for older releases and for projects that have not set up Trusted Publishing. It does not mean the release is malicious. It means you cannot check the origin cryptographically. The surrounding signals they name are maintainer history, release patterns, and repository activity. If you maintain the project, their suggestion is to use Trusted Publishing with a publisher that generates attestations on the next release.
Warnings are the other state, and they are not a polite version of “no badge.” PyPI says a warning is for metadata that looks unusual or that conflicts with the project’s history, including cases where a file may not have come from the maintainers. Account takeover and a taken-over publishing workflow are the examples they give. They also say many warnings have an ordinary explanation. The instruction is to check before you install, not to treat the banner as a conviction.
Four warning labels are specific enough to put in a review checklist.
Incomplete provenance: some files have attestations and some do not. Their example is wheels published through Trusted Publishing and a source tarball uploaded with an API token. They say to read the release notes for build trouble, or ask in the issue tracker, before installing.
Inconsistent provenance: files in the same release come from different build systems or different source repositories. Check the attestation metadata and see whether every file still comes from somewhere the maintainers control.
Change of provenance: this release’s files come from a different repository or workflow than an earlier release. An organization rename or a repository move will do that. So will a compromised pipeline. Their advice is to confirm a planned migration in the release notes or on a channel the project actually uses.
Loss of provenance: the project used to publish with attestations and this release does not. That can be a deliberate return to API tokens, or a config change that stopped generating attestations. If you were not told about the change, they say to treat the release with skepticism until someone verifies it.
The page chrome matches the labels. Provenance warnings sit at the top, next to banners for yanked releases, archived projects, and pre-releases. A quarantined project or release shows only the quarantine banner. That precedence is worth knowing so you do not go looking for a provenance badge that the quarantine view has replaced. When a release has full provenance or a warning, the sidebar also shows the state and links to the Security tab. Metadata on the project page is now ordered by usefulness and marked as “PyPI data” or “Data verified by PyPI,” instead of the older grouping by verification status alone.
Late files, and the July rule you can no longer break
The rebuilt release-files tab lets you search and filter a long file list, and it shows each file’s provenance state. Late files get an orange warning icon. PyPI defines late as a file uploaded more than 14 days after the release’s first timestamp. Those icons only appear on releases from before July 2026. That month, a policy started rejecting new files after 14 days. A late-file warning on an old release is a historical flag. It is not a path you can still use to append a wheel to a release from last Tuesday.
The release-history tab now shows file count, provenance state, and status: quarantined, yanked, archived, or pre-release. For a dependency review, that tab is often the faster read. A yanked release with full provenance is still yanked. Provenance does not outrank yank. Quarantine outranks the other banners. Status and origin are different columns, and the new history view finally prints both.
PyPI says the user docs grew alongside the tab. The package security guide, as they describe it on the blog, covers how to assess source, how to check provenance, how to verify integrity, how to protect yourself against common attacks, and how to report a problem. This note links the blog post rather than guessing a docs URL that was not in the fetched page. The work was funded by the Open Source Security Foundation and built by Kabu Creative. The blog thanks Dustin Ingram, Zach Steindler, the OpenSSF Securing Software Repositories Working Group, Mike Fiedler, Jacob Coffee, the PyPI team, and @cofiem for the release-files tab, plus the people who sat for interviews.
A case study of the warning patterns went into the OpenSSF style guide. The blog says that. It does not include a separate URL in the text reviewed here, so this note does not invent one.
What to do on a Tuesday install
A practical read of the tab, without adding rules PyPI did not write:
If the release is quarantined, stop. The page is telling you the quarantine matters more than the other badges.
If the release is yanked, do not pin it because the provenance looks tidy.
If provenance is full, you still know only what the attestation type covers. Trusted Publisher is a publishing path. SLSA build provenance is a commit and a builder. Neither one is a code review of the library’s behavior.
If provenance is missing, do not treat that as proof of malware, and do not treat it as proof of safety. Use the history and the repository, which is what PyPI says to do, and prefer a maintainer who has moved new releases onto Trusted Publishing if you have a choice.
If a warning is up, read it before the install finishes in CI. Incomplete, inconsistent, changed, and lost provenance are different stories. A planned repository move announced in the release notes is not the same event as a sudden loss of attestations with no note.
That is the limit of a user interface. It can show you a conflict. It cannot decide your exception process. Teams that want the tab to matter should fail a deploy, or at least page a human, when the state is a warning or a quarantine, and should record why they proceeded when they do. A tab nobody’s pipeline reads is a nicer project page and the same install habit.
Maintainers have the other half of the work. PyPI’s suggestion for a project with no provenance is Trusted Publishing through a publisher that emits attestations. Switching back to token uploads will, under the labels above, eventually read as loss of provenance to anyone watching the tab. If that switch is intentional, say so where the warning tells readers to look.
awswrangler 3.18.0 is a patch, not a badge
The Security tab will not catch CVE-2026-127119 for you. That bug is in the EMR support inside AWS SDK for pandas, the library published as awswrangler. AWS marked bulletin 2026-181-AWS as important. The publication line on the bulletin is 12 November 2026, 10:45 p.m. PST, with a last-updated line the same day. The advisory ID they print is GHSA-3rgm-255m-m9g5. They thank Upwind Security for coordinated disclosure.
Who has to move: versions 1.1.0 and later, up to but not including 3.18.0. The fix is 3.18.0. AWS also says forks and copied code need the same fix, not only the PyPI package you happen to import.
Impact, stated at the level the bulletin states it: cluster logs, including job step output, could be read by another party. That is enough to justify the upgrade. This note does not walk through how that read would be set up. The bulletin is the place for the vulnerability record. The action for operators is shorter.
Upgrade to 3.18.0. If you cannot upgrade today, pass an explicit log location in a bucket your account owns when you create clusters through this library. AWS also says you can create the default EMR log bucket in your own account, in each region you use. If that default name is already taken by another account, treat it as untrusted and use an explicit location you own. Do not keep writing logs at a name you do not control while you wait for a maintenance window.
After 3.18.0, callers who still rely on the default log location need s3:ListBucket on that bucket. AWS says the new version uses that permission to verify ownership. An upgrade that forgets the permission will fail the check the patch added. Put the permission in the same change as the version pin.
A 13 November post by Raajhesh Kannaa pointed at the same bulletin and the same two actions: move to 3.18.0, or pass your own bucket. A Japanese summary the same morning restated the affected range. Neither post adds a fact the bulletin does not already carry. They do show the advisory was circulating in public within a day of publication, which is a reason not to leave the pin for a quarterly batch if EMR logs from this library matter to you.
This is a different problem from the ansible-core 2.22 masking change covered in the 7 November note. That release was about keeping secrets out of automation logs you already own. This bulletin is about EMR logs landing where another account can read them. Masking a secret in a log you control does not help if the log bucket itself is not yours. Patch the library, and point the logs at a bucket the account owns.
A week that wants two habits, not one slogan
Supply-chain advice collapses, in bad write-ups, into “only install attested packages” or “patch everything on Tuesdays.” The documents from this week support neither slogan.
PyPI’s own post says missing provenance is common and is not a malware flag. A policy that blocks every unattested release will block a lot of ordinary history, including older pins you may still need. A policy that ignores warnings will miss the cases PyPI built the banners for: a file set that does not match, a publisher that changed, a project that dropped attestations without saying so.
The awswrangler bulletin is the opposite shape. The affected range is numeric, the fix version is numeric, and the temporary defense is “use a bucket you own.” There is no UI interpretation step. If awswrangler is in the environment and the version is below 3.18.0 and at least 1.1.0, and something in that environment creates EMR clusters, the bulletin applies. If you only import the library for S3 or Athena and never create clusters, read the bulletin before you decide the module split saves you. This note will not invent a claim that other modules are clean. AWS scoped the issue to the EMR module. Confirm that scope against the advisory if your import graph is unclear, and upgrade anyway if the pin is cheap.
Order the work like this. Pin awswrangler to 3.18.0 or newer where the library is installed, add the ListBucket permission if you use the default log location, and check forks. Then, on the next dependency review, open the Security tab for the packages you actually ship. Quarantine and yank stop the install. Warnings get a human. Missing provenance gets the history check PyPI described, not a fake green light and not an automatic ban you cannot staff.
The tab is a reading aid. The bulletin is a patch. Using one to excuse the other is how both get ignored.
Discussion
Leave a comment
No comments yet
Be the first to start the conversation.