groundy
Developer Tools

What PyPI's New Security Tab Proves, and What It Doesn't

PyPI's new security tab displays provenance states as a tiebreaker, not a gate. Vendor data shows attestations verify origin, not code safety, requiring manual review.

Published 7 references
On this page8 sections

Treat PyPI’s new provenance states as a tiebreaker, not an installation gate. According to PyPI’s vendor-stated November 6, 2026 announcement, full provenance means 100% of a release’s files carry attestations, and missing provenance does not necessarily indicate malice. The limit follows from what those attestations cover: the two supported types record the publisher, the source commit, the build platform and the build instructions, so provenance does not establish that the code is safe. The practical rule is to check the artifact, interpret warnings in context, and keep content review and registry status in the acceptance decision.

That distinction matters because the security tab makes provenance visible to people who may never have configured Trusted Publishing or read an attestation specification. A visible signal invites a binary interpretation: green means trustworthy, absent means suspicious. PyPI’s own documentation does not support either reading.

What shipped, and what the states describe

PyPI says its redesigned project pages are complete and live, with a security tab that communicates release provenance states and displays per-file attestations. PyPI calculates a release’s state by checking whether each release file has an attestation. This is therefore a statement about file coverage, not a score for the project as a whole.

Two attestation types are supported, and they answer slightly different questions:

  • PyPI Publish attests that a file was published by a PyPI Trusted Publisher.
  • SLSA Build Provenance links a file to its exact source commit, build platform, and build instructions.

A release labeled Full Provenance has attestations for 100% of its files (PyPI’s announcement). That does not necessarily mean every file carries both types or supplies the same source and build detail. The label certifies coverage; the individual attestations determine which origin claims you can inspect.

The remaining signals are No Provenance and four warnings: Incomplete, Inconsistent, Change of Provenance, and Loss of Provenance. All of these names and behaviors come from PyPI’s announcement, rather than an independent audit of the verification system.

An attestation proves its signed claims, not everything about the package

General attestation architecture helps explain the boundary. A study of attestation mechanisms describes claims being collected and cryptographically signed to form evidence, which a verifier can then assert or deny. The evidence establishes the claims it contains; it does not acquire additional guarantees merely because the signature verifies (arXiv:2204.06790). That paper concerns attestation mechanisms broadly, not PyPI’s implementation.

Applied to PyPI, a Publish attestation supports a claim about the publisher. A SLSA Build Provenance attestation supports claims about source and construction. Neither type, as PyPI’s documentation describes them, claims that the source is benign, the maintainer has good intentions, or the package’s dependencies are safe.

A second paper separates attestation into three roles: attester, verifier, and relying party. The verifier appraises the evidence, but the relying party makes the trust decision (arXiv:2312.08903). On the new project page, PyPI computes and displays the state. The person or platform accepting the dependency remains the relying party.

That is the operational consequence of the redesign. PyPI can make origin evidence easier to find, but it cannot turn that evidence into a complete installation decision. The tab reduces the effort required to locate provenance information; it does not remove the work of appraising risk.

What full provenance leaves open

The largest residual risks concern what code does, rather than where a file came from.

The DySec preprint reports that the Pymafka package activated a keylogger immediately after installation, bypassed sandbox checks, and stole sensitive device data. It also reports that detection methods struggle with multiphase threats hidden in nested dependency chains (arXiv:2503.00324). An origin record for the top-level artifact does not certify the behavior executed during installation or the contents of an indirect dependency chain.

The same study identified six packages previously classified as benign that exhibited malicious installation behavior, including data exfiltration, port scanning, and unauthorized remote access; four were subsequently removed by PyPI maintainers. That result is preprint-reported and predates the security tab. It does not show that these packages had full provenance. It does show why an origin label cannot substitute for examining behavior.

Lookalike and replicated packages present a different problem. An FSE 2026 study (viewed via its arXiv version) reports sampling approximately one-third of the 670,000 packages available on PyPI as of August 2025 (arXiv:2606.29785), a candidate dataset of 200,737 packages drawn as the pool for replication detection. Within that pool it identified 1,361 replicated packages (arXiv:2606.29785); those are similarity findings, not a count of confirmed malware. The researchers also note that users rely on package names and versions when identifying packages to install. Provenance does not check whether a familiar-looking name belongs to the project you intended to select.

PyPI’s own installation documentation preserves the same distinction: installing a package means running third-party code, with inherent security risks, and it recommends consulting the package security guide before installation (PyPI user documentation). Full provenance can answer “where did this file come from?” while leaving “what will installing it do?” untouched.

A warning is a reason to investigate, not a compromise verdict

The four warnings describe different kinds of mismatch:

  • Incomplete: some files are attested and others are not. PyPI’s example is wheels published through Trusted Publishing while the source tarball is uploaded using an API token.
  • Inconsistent: files within one release come from different build systems or source repositories.
  • Change of Provenance: the repository or workflow differs from a previous release.
  • Loss of Provenance: a project that previously supplied attestations uploads files without them.

PyPI says these warnings are intended to help identify files that may not have come from the project’s maintainers, including cases involving account or publishing-workflow takeover. It also says many warnings have legitimate explanations, such as organization renames, repository migrations, or switching back to API tokens (PyPI’s announcement).

The same signal can therefore point toward either an ordinary publishing change or a serious event. Treating every warning as compromise would erase that distinction. Treating every warning as routine would erase its value. The useful response is to ask whether the project’s history and release pattern explain the change, and to escalate when they do not.

A reviewer’s trust map

The recommended actions below are an inference from PyPI’s state definitions and fallback guidance, not the measured outcome of a tested dependency policy.

SignalWhat it establishesWhat remains openRecommended response
Full ProvenanceEvery release file has an attestation. Its type determines the available publisher, source, and build claims.Code behavior, maintainer intent, package identity, and nested dependencies.Count it as positive origin evidence. Continue content and dependency checks; use it as a tiebreaker between otherwise comparable candidates.
No ProvenanceNo file in the release has an attestation, so file origin cannot be cryptographically verified through this mechanism.Whether the absence reflects age, publishing setup, or something else.Do not reject on absence alone. Check maintainer history, release patterns, and repository activity.
IncompleteAttestation coverage differs across files in the release.Why particular artifacts lack coverage and whether your selected artifact is attested.Inspect the per-file attestations and identify the publishing difference. Do not assume the whole release shares one file’s evidence.
InconsistentFiles in one release have different build-system or source-repository origins.Whether that difference is expected for the project.Look for a release or repository explanation. Pause a high-consequence update if the difference is unexpected and unresolved.
Change of ProvenanceThe repository or workflow differs from a previous release.Whether the change is a migration, rename, workflow edit, or takeover.Compare project history and prior releases. Verify an unexplained change before treating it as routine.
Loss of ProvenanceA previously attesting project has uploaded unattested files.Why attestation stopped and whether the publishing process changed.Investigate the workflow change and surrounding trust signals. Absence after prior coverage deserves attention, but is not itself proof of malice.

Other release banners answer other questions and should not be folded into the provenance score. PyPI says late-file warnings identify files uploaded more than 14 days after the initial release timestamp. Because a July 2026 policy rejects new files after 14 days, these warnings appear only on releases from before that policy. Quarantine has stronger display precedence: when a project or release is quarantined, PyPI shows only the quarantine banner, ahead of provenance warnings and yanked, archived, or pre-release banners (PyPI’s announcement). An acceptance workflow should check those registry statuses separately rather than treating provenance as a summary of them.

Why absence makes a poor gate

PyPI explicitly says No Provenance is typical for older releases and projects that have not configured Trusted Publishing. It directs users to maintainer history, release patterns, and repository activity when provenance is absent. In other words, the vendor’s prescribed fallback is appraisal, not automatic rejection.

A strict gate would consequently reject legitimate releases because of their age or publishing configuration. None of the sources cited here quantifies attestation adoption, so this article cannot estimate how many packages such a gate would affect.

Independent detection research supplies a related caution. An ACSAC 2023 study of malicious-package detection across npm and PyPI reports that no ground truth exists for benign packages and that some packages labeled benign could still contain malicious code. Citing follow-up work, it notes that repositories need detectors with very low false-positive rates (arXiv:2310.09571). That study does not measure PyPI’s provenance warnings or calculate their false-positive rate. Its relevance is narrower: converting an incomplete signal into an accusation is a poor design pattern for repository security.

Surrounding signals are not guarantees either. Repository activity and maintainer history can provide context for a decision, but they do not establish artifact equivalence, safe contents, or benign intent. They are the next evidence to inspect, not a replacement certificate.

A dependency policy that keeps provenance in proportion

I would write the acceptance rule in this order:

  1. Apply the actual gates first. Check quarantine and yank status, then perform the content review appropriate to the dependency, including install-time behavior and transitive dependencies. Full provenance should not waive these checks.
  2. Inspect provenance at the artifact level. Determine whether the file you intend to install is attested and which attestation type it carries. Do not transfer a wheel’s evidence to an unattested source tarball, or assume every attestation contains SLSA source and build claims.
  3. Use full provenance as positive, bounded evidence. When candidates are otherwise comparable, prefer the one with clearer origin evidence. That preference does not make it safe or interchangeable with another version.
  4. Use the prescribed fallback when provenance is absent. Review maintainer history, release patterns, and repository activity instead of treating the package as suspect by default.
  5. Require an explanation for consequential anomalies. For a high-impact dependency update, an unexpected Change of Provenance or Loss of Provenance is a sensible reason to pause and verify through the project’s available history and communication channels. The pause follows from uncertainty; it is not a declaration of compromise.

This is deliberately less satisfying than a universal ‘attested or reject’ rule. It requires judgment, and some investigations will end with an ordinary migration rather than a security finding. But that work reflects what the evidence can actually establish.

What remains unverified

The trust boundary here is assembled from vendor-stated UI semantics, general attestation literature, and PyPI threat studies that predate the security tab. None of the cited studies reports independent replication of its findings. No source cited here independently audits PyPI’s attestation verification path, quantifies what share of projects attest, evaluates the tab’s effect on installation decisions, or tests the policy recommended above.

The evidence also does not settle whether a compromised maintainer account or seized publishing workflow can emit internally consistent attestations. Nor do the cited threat studies demonstrate that install-time malware, nested attacks, or replicated packages have co-occurred with a Full Provenance label. Those possibilities should not be reported as observed outcomes.

Until that evidence exists, the defensible reading is narrow: provenance can verify specified origin claims for release files. It cannot certify code, intent, or dependency behavior, and its absence is not evidence of malice.

References

Follow the links in the article for context. The supporting material is collected here for further reading.

  1. PyPI's vendor-stated November 6, 2026 announcementblog.pypi.orgAccessed
  2. arXiv:2204.06790arxiv.orgAccessed
  3. arXiv:2312.08903arxiv.orgAccessed
  4. arXiv:2503.00324arxiv.orgAccessed
  5. arXiv:2606.29785arxiv.orgAccessed
  6. PyPI user documentationdocs.pypi.orgAccessed
  7. arXiv:2310.09571arxiv.orgAccessed

Join the discussion

Share a useful perspective or ask a question about this article.

Discussion guidelinesComments privacy