PSF says pip before 26.2.2 bundled compromised eMudhra certificates; its CVE came 31 days after the pip fix
CVE-2026-133949, published on 20 November 2026, says pip before 26.2.2 bundled certifi with compromised eMudhra certificates. pip 26.2.2 is dated 20 October, 31 days earlier (derived), and no saved page says why the record came later.
By Parminder Kumar Sharma · LinkedIn · · 17 min read
31 days between the pip fix and the pip CVE (derived)
pip's own changelog dates pip 26.2.2 as 2026-10-20 and lists one change: "Upgrade certifi to 2026.10.19". The CVE record for pip, CVE-2026-133949, was published at 19:12 UTC on Friday 20 November 2026. The gap is 31 days (derived: 11 days to 31 October plus 20 days of November). The record, assigned by the Python Software Foundation (PSF), says: "pip prior to v26.2.2 the certifi bundled with pip contains the compromised eMudhra certificates." PSF's announcement opens: "There is a CRITICAL severity vulnerability affecting pip." The score on the record is CVSS 4.0 9.3, Critical, assigned by PSF.
What that does not establish. It does not show that anyone was unprotected for 31 days: the fixed pip existed on 20 October, and the Python Discourse topic that opens at 16:57 UTC that day already names 26.2.2 as the minimum. It does not say why the record came later: no saved page does, and this briefing does not guess. It does not show exploitation: the CISA-ADP entry on the record, time stamped 19:20 UTC on 20 November, reads Exploitation none, Automatable yes, Technical Impact total, which is the coordinator's classification at one moment and not a finding. It does not say how many installations run a pip below 26.2.2, and it does not say whether pip releases after 26.2.2 change anything: pip 26.3 of 2 November lists "Upgrade certifi to 2026.10.23", and the changelog gives no reason for it.
How this was read. Every page was read on 21 November 2026 from copies this site saved between about 08:42 and 08:50 UTC. Several are living pages and are quoted as they stood then: the CVE record was last updated at 19:59 UTC on 20 November; GitHub's copy of the pip advisory is marked Unreviewed; the Discourse topic's opening post was last edited on 31 October at 18:58 UTC, and the saved copy holds 20 of the topic's 61 posts. The CVE is PSF's own account of a flaw in software its ecosystem ships, so it is a primary for what the record says and not an independent audit. No individual is named here: the pages credit people, and only organisations appear. Every figure this site worked out is labelled derived.
What happened, in the order the pages give it
The incident tracker says the keys of eight eMudhra intermediate certificates were shown to be compromised and that the report reached the root programs on 17 October. It dates the blocks: Firefox and Chrome from 17 October, Microsoft platforms from 20 October, Apple products from 22 October. On Sunday 18 October the pip pull request 14371 was merged, upgrading pip's vendored certifi to 2026.10.18 on the main branch; the page shows 45 of 46 checks passed and does not say which did not. The pull request description carries the line "Assisted-by: Claude", under pip's AI policy checklist; the page does not say what the tool did.
On 20 October three things landed. certifi's advisory was published, Critical, 9.1. pip 26.2.2 was released with certifi 2026.10.19. And the Discourse topic opened with a list of 18 entries of minimum versions and the warning that failure to upgrade "could result in monster-in-the-middle (MITM) attacks against TLS clients, including python.org and pypi.org". The next day a CPython issue proposed a denylist in Python's ssl module; PSF published CVE-2026-115557 for it on 24 October at 14:15 UTC, 27 days (derived) before the pip record. The diagram puts every date on one scale and zooms on the crowded week. Events the pages date by day only are drawn as whole day bars, not as points.
Later dated entries: on 2 November pip 26.3 upgraded certifi again, to 2026.10.23 (13 days after 26.2.2, derived); on 10 November the tracker added that Apple fixed its developer tools in macOS 27.2.1, 26.7.3 and 15.8.3, and the Python CVE record was updated. Then the pip record, 33 days (derived) after the pull request and 34 days (derived) after the report to the root programs.
What the pages say about the keys, and what they leave open
The saved pages give the cause in plain words and print no key value, and this briefing prints none. The incident tracker says: "The private keys for the intermediate certificates were incorrectly generated by the CA." and "This means that anyone can recover them from the certificates." A participant in the Python thread calls it "a bad private key". certifi's advisory puts it as a fact about the bundle: "Eight eMudhra intermediate CA certificates that chain to this root have compromised private keys and the CA has revoked them."
The tracker says the certificates were created in April 2025, about 18 months before the report (derived), and that browsers could not have been caught: "CT-enforcing clients like Firefox, Chrome, Edge and Safari could not have been attacked". It then names the gap that matters for pip: "We do not know if attackers exploited the flawed key generation to target these clients." The clients at risk are the ones outside the browser ecosystem, and the tracker's explanation is that "many other clients do not consume revocation information and require software updates".
Two earlier briefings on this site cover other parts of the same incident. Brief 481 reports Microsoft saying it removed six eMudhra G3 roots and an AAA Certificate Services cross sign from the Windows trust list on 26 October, in an update that only the Microsoft Update Catalog offers, after updating its Disallowed list on 20 October. Brief 550 reports Cloudflare's post of 30 October: about 22,000 attempted handshakes with the weak keys in its own outbound traffic, "no evidence of the weak keys being used to intercept Cloudflare's connections so far", and more than half of its probes on large networks still accepting the chain. Neither is evidence about pip, and neither shows exploitation against a victim.
Why pip needed its own fix, and where a fix can sit
An operating system update can fix most software, because most software asks the operating system whom to trust. pip does not rely on that alone. A participant in the Python thread says pip "always loads the CA certificates from its vendored certifi in addition to the system trust store, and there is no supported way to disable this", and concludes: "Updating pip is the only reliable fix." The participant links a pip issue titled "Switch to using truststore only by default (currently mixed certifi + truststore)"; this site did not read it. pip's changelog agrees on the mixture: since 24.2 it uses "system certificates and certifi certificates" on Python 3.10 and later, and on 3.9 and earlier only certifi.
That is why the CVE speaks of Python and not of the operating system. Its condition is "pip is not vulnerable when using a Python version that rejects the compromised certificates." PSF's other record, CVE-2026-115557, describes the Python side as a "denylist for known compromised certificates" in the ssl module, and the issue that proposed it says the check runs when verification is enabled and that CERT_NONE, which switches verification off, is unaffected. A Python participant describes it as "a defense in depth that prevents connections when the OS, certifi, or pip are not up to date".
Where each saved page says a fix can act on pip (sources as named in each cell, read on 21 November 2026)
| Layer | What the saved pages say | Source |
|---|---|---|
| The operating system's trust store | Does not reach pip's vendored certifi, a participant says. A participant believes Debian patches its packaged pip to use the system store; the same thread notes Debian's python-pip was at 26.1.2. | Python Discourse topic |
| pip itself | pip 26.2.2 vendors certifi 2026.10.19. The CVE counts every pip below 26.2.2 as affected. | pip changelog, CVE record |
| certifi, in other packages | Upgrade to 2026.10.19 or later and rebuild container images that vendor certifi. | certifi advisory |
| Python's ssl module | A Python that rejects the compromised intermediates makes pip not vulnerable, the CVE says. Fixed in 3.11.18, 3.12.16, 3.13.18, 3.14.10 and 3.15.2. | CVE records |
| macOS built in pip3 | The tracker says the built in pip3, and pip in virtual environments made from the built in python3, use their own outdated copy of the roots. Apple's fix of 10 November is listed in macOS 27.2.1, 26.7.3 and 15.8.3. | Incident tracker |
One incident, three records, three scores
The three records that touch pip score the same incident at 9.1, 9.3 and 9.4. All three are CVSS 4.0, all three are Critical. The numbers differ because the vectors differ, and the pages do not explain why (derived by comparing the strings metric by metric).
The three scores, who the page says assigned them, and the metrics that differ (derived by comparing the printed vectors)
| Record | Score as printed | What differs from the pip record |
|---|---|---|
| certifi advisory GHSA-6q48-6c7w-9gw2, 20 Oct | Critical 9.1 (shown on GitHub's page; the page does not say who scored it) | Attack Complexity High and Attack Requirements Present, where the pip record has Low and None |
| pip CVE-2026-133949, 20 Nov | Critical 9.3 (PSF, in the CVE record) | The reference for this comparison |
| CPython CVE-2026-115557, 24 Oct | Critical 9.4 (PSF, in the CVE record) | User Interaction Passive, and High impact on availability, plus all three impacts High on the subsequent system |
The CISA-ADP entries differ too (SSVC entries from the CISA coordinator role): Exploitation none on both, Technical Impact total on both, but Automatable yes on the pip record and no on the Python record. This site does not say which is right, and no page says why they differ. The reading that fits the pages is that the records score different artefacts of one incident: a bundle, a vendored copy and an interpreter's missing check (this site's inference). A reader who sorts by score should treat the three as one exposure, not three separate ones (this site's suggestion).
certifi advisory, 9.1: CVSS:4.0/AV:N/AC:H/AT:P/PR:N/UI:N/VC:H/VI:H/VA:N/SC:N/SI:N/SA:N
pip CVE-2026-133949, 9.3: CVSS:4.0/AV:N/AC:L/AT:N/PR:N/UI:N/VC:H/VI:H/VA:N/SC:N/SI:N/SA:N
CPython CVE-2026-115557, 9.4: CVSS:4.0/AV:N/AC:L/AT:N/PR:N/UI:P/VC:H/VI:H/VA:H/SC:H/SI:H/SA:H
product: pip versions: version 0, lessThan 26.2.2 (versionType python) defaultStatus: unaffected
What Python's denylist covers, and what it does not
The pip record leans on Python. Its first route is to upgrade Python "to a version that rejects the compromised intermediates (CVE-2026-115557)". The Python record lists the affected and fixed lines as follows.
CVE-2026-115557 as the record prints its table (version fields as printed, affected module ssl, read 21 November 2026)
| Python line | Affected, as printed | Fixed in |
|---|---|---|
| 3.11 and older | version 0 up to, not including, 3.11.18 | 3.11.18 |
| 3.12 | 3.12.0 up to, not including, 3.12.16 | 3.12.16 |
| 3.13 | 3.13.0 up to, not including, 3.13.18 | 3.13.18 |
| 3.14 | 3.14.0 up to, not including, 3.14.10 | 3.14.10 |
| 3.15 | 3.15.0 up to, not including, 3.15.2 | 3.15.2 |
| 3.16 | 3.16.0a1 up to, not including, 3.16.0 | 3.16.0 |
Three limits are on the pages. First, the list is of known certificates: PSF's announcement says compromised certificates "on this denylist will automatically be rejected", and the CPython issue sets the rule for adding more: "The block list is designed as last line of defense." A new compromise therefore needs a new Python release (this site's reading). Second, the project argued about it. One participant objected that "Updating a Python release is nowhere near as safe." Another worried about "signing us up for more expedited releases and asking users to mitigate by updating the Python runtime". The issue's author replied that packages shipping their own trust stores are the problem, and that the author knew of one with almost 14 million downloads per day that ships eMudhra certificates; the page names none. Third, the end of life lines. The thread's opening post says Python 3.10 and older are "end-of-life and no longer eligible for security fixes", yet the record's first row starts at version 0. The issue's timeline lists commit titles that mention 3.10 and 3.9 version strings and a 3.10 backport dated 7 November; the saved page does not say whether any of them is an official release.
For those interpreters the pages leave a gap. pip's changelog says 26.1 (2026-04-26) dropped support for Python 3.9, and a participant writes that 26.2.2 does not support environments older than 3.10, adding "I suspect those are not at risk?" No saved post answers it. Whether a Python 3.9 environment is exposed, and by what route it could be fixed, is not stated on any saved page (the saved copy holds 20 of 61 posts, so a later reply may exist).
Installing the fixed pip without trusting the old one
A participant asked the obvious question: how to update pip "without running an older version of Pip that's vulnerable to MitM attacks". The reply, and the opening post, give a command that installs the 26.2.2 wheel from a direct URL with its SHA 256 checksum in the URL, so the check does not depend on certifi: from that install, the reply says, "you can bootstrap into a secure installation of certifi, uv, or other dependencies". The CVE's own words are "use hash pinning with a known-secure version of pip". This briefing prints neither the address nor the checksum: copy them from the Discourse topic itself and not from a second-hand copy; checking them against what pip's project publishes is this site's suggestion.
For uv, a participant describes updating it using the host's CA bundle instead of the embedded one, on the stated assumption that the system certificates are already updated; on Linux that depends on the distribution. For virtual environments, the thread offers a user's script that finds them and upgrades pip from the checked wheel; this site did not run or read it beyond the saved text, and it skips environments older than 3.10.
The ensurepip question was asked on the day: "Is anything needed to update the pip shipped with Python as part of the ensurepip module?" The reply: "GH-159910 is merged and has backports open, with plans to cut security releases as soon as reasonably possible." It adds that an enterprising user can patch an existing installation by hand; this site does not print the steps. The saved pages do not print which pip version each of 3.11.18 to 3.15.2 bundles, so the pip inside a new virtual environment from one of them is a version this site cannot state (not stated). The thread's workaround for the interim was upgrading pip after creating any new environment. Two other places carried old pips on 21 October: a participant found the Docker official Python images still shipping pip 26.2.1, and a reply says those images are maintained downstream of CPython and "I expect they will update once our updated releases are out". The opening post, as last edited on 31 October, lists the Docker official Python images for the five fixed Python versions. One more trap from the thread: "if you use dependency cooldowns, you'll be vulnerable till you update".
What the pip record states, and what it does not
The pip record and its neighbours: stated, and not stated (as read on 21 November 2026)
| Topic | Stated | Not stated |
|---|---|---|
| Affected versions | CVE record: version 0 up to, not including, 26.2.2. GitHub's copy of the advisory shows Affected versions Unknown and Patched versions Unknown, and no package. | A first affected version above 0, or the certifi version inside each pip |
| The fix | 26.2.2, with certifi 2026.10.19 per pip's changelog | The record does not name 2026.10.19; it says only prior to v26.2.2 |
| The Python condition | Not vulnerable on a Python that rejects the compromised certificates | The Python version numbers: the record points to CVE-2026-115557 instead |
| Impact | CWE-295 improper certificate validation; confidentiality and integrity High | Any attack scenario, victim or method; discovery is recorded as UNKNOWN |
| Exploitation | CISA-ADP, 19:20 UTC on 20 November: Exploitation none | Any evidence beyond that classification |
| Timing | Reserved at 17:03 UTC, published at 19:12 UTC on 20 November: 2 hours 9 minutes (derived) | Why 20 November |
| The announcement | PSF's list message of 7:08 p.m. on 20 November: "CRITICAL severity" | The time zone. If it were UTC the message would precede the record by 4 minutes 43 seconds (derived), so this site does not rely on it |
| Reach | Any pip below 26.2.2 is within the record | How many installations run one; whether any index or mirror was targeted |
Where the pages differ, left unresolved
The patched certifi. certifi's advisory lists 2026.10.18 as the patched version and says upgrade to 2026.10.19 or later. The thread's footnote explains: "2026.10.19 is recommended for compatibility with unaffected emSign root certificates". pip's pull request put 2026.10.18 on the main branch on 18 October; the release of 20 October carries 2026.10.19. The advisory's patched version and the minimum everyone is told to install are different numbers, and the later pip 26.3 carries 2026.10.23, whose content no saved page describes.
uv. The thread's list says uv 0.13.6 and marks it as a package that did not bundle the roots; the tracker lists uv 0.13.5 as patched. Neither page says which a reader should hold to.
The two changelogs. pip's changelog page, as saved, contains neither eMudhra, emSign nor either CVE id: the 26.2.2 entry is one line about a vendored library. A reader who follows only release notes would not learn that the release was a security fix. The CVE and the announcement came after.
OS level against interpreter level. A participant says distrusting the intermediates at the operating system level does not help pip below 26.2.2. The CVE says pip is not vulnerable on a Python with the denylist. These are about different layers and do not contradict each other (this site's reading); the pages do not set them side by side.
What to do, in order
What PSF and the Python thread say first: the record's two routes are to upgrade Python to a version that rejects the compromised intermediates, or to use hash pinning with a known-secure pip; the thread asks for the minimum versions in its list and for container images to be rebuilt with layer caching disabled. The rest are this site's suggestions.
Take this with you
In order of value (items 4 to 7 are this site's suggestions)
- Python, per PSF: move to 3.11.18, 3.12.16, 3.13.18, 3.14.10 or 3.15.2 or later, so the ssl denylist is present.
- pip, per PSF and the thread: install at least 26.2.2 by a route that does not trust the old pip, using the thread's hash pinned direct URL command copied from the thread.
- certifi, per the advisory and the thread: 2026.10.19 or later wherever it is installed or vendored, and rebuild container images with layer caching disabled.
- Scan base images, virtual environments, CI runners and build caches for pip below 26.2.2 and certifi below 2026.10.19, because an environment made last month keeps its old pip until someone upgrades it.
- Work through the thread's list of tools (build front ends, installers, publishing actions, Python distributions) and record the version each of your pipelines uses.
- Python 3.10 and older: plan the move, because the saved pages do not establish an official fix route for them, and the CVE's not vulnerable condition cannot apply to a Python without the denylist.
- Update operating systems and distribution packages too (the tracker lists macOS 27.2.1, 26.7.3 and 15.8.3), and check built in pip3 on macOS separately.
The question the 31 days leave open
pip's fix reached users on 20 October and its CVE record on 20 November. Which of your pipelines would have noticed a pip below 26.2.2 on 21 October without a CVE, and does anything in your tooling wait for a record before it treats a vendor's fix as a fix?
Key facts
Sources
- PrimaryCVE-2026-133949, pip bundles compromised eMudhra certificates via certifi, published 20 November 2026 at 19:12 UTC, updated 19:59 UTC, with the CISA-ADP SSVC entry of 19:20 UTC, read on 21 November 2026CVE.org record (assigner PSF)accessed 2026-11-21
- PrimaryAnnouncement of CVE-2026-133949, 20 November 2026, 7:08 p.m. as the list page prints it with no time zone: CRITICAL severity, read on 21 November 2026Python security-announce mailing listaccessed 2026-11-21
- Primarycertifi bundles emSign root CAs with compromised intermediate CA keys, published 20 October 2026, Critical 9.1: patched version 2026.10.18, upgrade to 2026.10.19 or latercertifi advisory on GitHub (GHSA-6q48-6c7w-9gw2)accessed 2026-11-21
- PrimaryUpgrade certifi to 2026.10.18, merged on 18 October 2026, 45 of 46 checks passedpip pull request 14371accessed 2026-11-21
- Primarypip 26.2.2 (2026-10-20): Upgrade certifi to 2026.10.19; 26.3 (2026-11-02): Upgrade certifi to 2026.10.23; 26.1 dropped Python 3.9; 24.2 certificate behaviour; read on 21 November 2026pip changelogaccessed 2026-11-21
- PrimaryCVE-2026-115557, Distrust compromised intermediate certificates in ssl module, published 24 October 2026, updated 10 November: affected and fixed Python versions, CVSS 4.0 9.4, CISA-ADP SSVC of 2 NovemberCVE.org record (assigner PSF)accessed 2026-11-21
- PrimaryAnnouncement of CVE-2026-115557, 24 October 2026, 2:20 p.m. as the list page prints it: a denylist of known compromised certificates in the ssl modulePython security-announce mailing listaccessed 2026-11-21
- Primaryssl: distrust compromised eMudhra intermediate CA keys, opened 21 October 2026, closed 25 October: the proposal, the debate on expedited releases and the denylist policyCPython issue 160040accessed 2026-11-21
- PrimaryInformation for Python users regarding eMudhra / emSign CA certificate security incident: opening post of 20 October 2026 16:57 UTC (last edited 31 October) with the minimum versions list, and 19 replies, read as JSON saved on 21 November 2026Python Discourseaccessed 2026-11-21
- Reported byDatabase entry for CVE-2026-133949, published 20 November 2026, Critical 9.3, marked Unreviewed, with Affected and Patched versions shown as UnknownGitHub Advisory Database (GHSA-jg88-vh4g-v698)accessed 2026-11-21
- Reported byIncident tracker dated 20 October 2026, updated 10 November: the cause of the keys, the block dates, who has patched, workarounds; a community resource, not a vendor statementeMudhra Intermediate CA Compromise incident trackeraccessed 2026-11-21


