On 20 November a Hacker News front-page post was titled "OpenSSL: Buffer overflow in EVP_EncodeBlock() (CVE-2026-121481, CVSS 9.8)". A day later, NVD showed 5.9. OpenSSL's advisory said Low. The GitHub Advisory Database still showed 9.8.
None of those sources is lying. The same bug was given three different ratings by three different parties inside two days, and the dispute was still open when we read it on 21 November. This post sets out what the bug is, where each number came from, and which one should decide when you patch.
What the bug is
OpenSSL is the CNA for its own CVEs, so the advisory text in NVD is OpenSSL's own. Its summary:
"Issue summary: Calling EVP_EncodeBlock() with an input length close to INT_MAX may overflow the calculation of the encoded output size and write past the end of the caller-supplied output buffer."
EVP_EncodeBlock() base64-encodes a buffer. The caller passes the input length and supplies the output buffer. The arithmetic that computes the output size is done in a signed int, so when the input length is close to INT_MAX the result wraps and the function writes past the end of the buffer. That is a heap buffer overflow in the process that made the call (CWE-190, integer overflow or wraparound).
The condition is narrow. In OpenSSL's words:
"The severity is Low because no attacker-controlled path to the overflow is known. The caller must already hold a buffer of about two gigabytes before the defect can be reached."
The advisory also says which parts of OpenSSL are not a path:
"OpenSSL's TLS, QUIC and CMS implementations never encode buffers of that size, and the command line tools cannot practically be made to do so: the limits on reading a proxy password from a file and on bignum sizes prevent it. An application that accepts an attacker-controlled length and forwards it to EVP_EncodeBlock() without its own bound check is exposed."
So the exposed party is not "anything that links OpenSSL". It is an application that calls EVP_EncodeBlock() itself, with a length an attacker can push near 2 GB, and without checking that length first. The advisory also states: "The affected code is outside the FIPS module boundary."
Where each number came from
NVD keeps a change history for every CVE record. Read together with the public CISA issue thread, it gives this sequence:
- 18 November. OpenSSL publishes the CVE with a one-line description.
- 19 November. CISA, which adds its own scores to NVD records, scores it
AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H. That vector is 9.8: network-reachable, low complexity, full loss of confidentiality, integrity and availability. - 20 November, early UTC. The Hacker News post goes up with 9.8 in its title. At that moment, 9.8 was the score on the NVD record.
- 20 November. OpenSSL replaces the one-line description with the full advisory quoted above. A member of the OpenSSL security team opens an issue on CISA's vulnrichment repository saying both of the CVEs it names "were scored while the records only had a one-line summary", and asking CISA to take another look. NVD's change history bears out that order: the 9.8 was added on 19 November, while the record still carried the one-line description, and OpenSSL's full text arrived about twelve hours later.
- 20 November. CISA changes the vector twice: first to a local attack vector, then to
AV:N/AC:H/PR:N/UI:N/S:U/C:N/I:N/A:H. That is 5.9, Medium, with no confidentiality or integrity impact, only availability. It is the score NVD shows now, marked as a secondary score from CISA. - 20 November. The researcher who says they reported the bug replies on the same thread and argues for 9.8. Their case is that CVSS scores a library for the reasonable worst-case application that uses it, and that a near-identical overflow in another crypto library API, CVE-2022-37454, was scored 9.8 by both NVD and CISA.
The 9.8 that lasted about 21 hours on the NVD record had not gone everywhere else when we read on 21 November. The GitHub advisory for this CVE (GHSA-83xj-7xh4-4p7x) still showed CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H, 9.8. A scanner that reads one feed and a scanner that reads another can give you different numbers for this CVE on the same day.
CISA's other assessment on the record, its SSVC entry, moved with the score. The entry added alongside the 9.8 read exploitation: none, automatable: yes, technicalImpact: total. The entry that replaced it alongside the 5.9 reads exploitation: none, automatable: no, technicalImpact: partial. Only the exploitation field was the same in both.
Why the vendor and the score can both be "right"
OpenSSL's security policy says it does not use CVSS to set severity. It gives a reason: CVSS "does not properly reflect the broad usage of the OpenSSL Library" and does not account for how likely a configuration is to be affected. The policy warns that scores from independent parties may differ sharply from OpenSSL's own rating.
The two scales answer different questions:
- A CVSS base score asks how bad things could be for the worst plausible user of the component, assuming the vulnerable setup is present. For a library, that is an imagined application that passes attacker-controlled lengths straight through.
- A vendor severity like OpenSSL's asks how likely that setup is across real deployments. For this bug, OpenSSL found no path through its own protocols or tools.
The reporter's 9.8 and OpenSSL's Low can therefore both follow their own rules. The middle number is the one the thread was still arguing over. Whichever way it settles, the advisory text does not change, and that text is what tells you whether you are affected.
What to actually check
- Do you call
EVP_EncodeBlock()directly? If you write C or C++ against OpenSSL, search your own code:grep -rn EVP_EncodeBlock .For each call, find where the length comes from. If an outside party can set it and nothing caps it well belowINT_MAX, add a bound check now. That fix is yours to make and does not wait for an OpenSSL release. - Read the vector, not the number.
C:N/I:N/A:HandC:H/I:H/A:Hlead to very different responses. If a ticket or a dashboard only quotes a score, open the record and read the vector and the advisory text before you book an emergency window. - Write down which feed your scanner uses. If your policy says "patch anything at 9.0 or above within 24 hours", this CVE crossed that line in one feed and not in another on the same day. The policy is only as stable as its source.
- Take the fix through your normal update path. The pull request carrying the fix (openssl/openssl#33652) was still open when we read it, and we did not find a released OpenSSL version that includes it. Under its policy, OpenSSL ships Low-severity fixes in its next scheduled patch release. If you use distribution packages, watch the package changelog for the CVE ID rather than the upstream version number.
Treat this one as a library bug that needs a specific mistake in the calling code before it can be reached. It does not need a weekend outage. A number in a headline is a reasonable cue to go and read the advisory. It is not a substitute for reading it.
Sources, read 21 November 2026: the NVD record for CVE-2026-121481 and its change history (NVD API); cisagov/vulnrichment issue #354; the GitHub advisory GHSA-83xj-7xh4-4p7x; openssl/openssl pull request #33652; the OpenSSL security policy; and the Hacker News submission of 20 November.