Skip to main content

The OpenSSL bug at the top of Hacker News today carries a headline score of CVSS 9.8: “Buffer overflow in EVP_EncodeBlock() (CVE-2026-121481)”. OpenSSL, which issued the CVE and is fixing the bug, rates it Low. Both ratings sit in the same public CVE record. The 9.8 is not OpenSSL’s. It came from CISA’s Vulnrichment programme a day after publication, and on NVD it is the only number on the page.

That looked like a one-off, so we checked whether it was. We pulled the CVE record for every OpenSSL vulnerability with a 2023 or later CVE ID and compared the vendor’s severity with CISA’s score. It is not a one-off. It is how the system works, and if your patching policy runs on a number it is quietly making decisions for you.

TL;DR

  • CVE-2026-121481 is rated Low by OpenSSL and 9.8 Critical by CISA’s ADP (Authorised Data Publisher) container. OpenSSL says no attacker-controlled path to it is known and the caller must already hold a buffer of about 2 GB.
  • Across 103 OpenSSL CVEs with 2023-or-later IDs, CISA’s CVSS band was higher than OpenSSL’s severity 88 times, the same 15 times, and lower zero times.
  • Of the 78 that OpenSSL rated Low, CISA scored 72 as Medium or worse, 35 at 7.0 or above, and 6 as Critical.
  • CISA’s own SSVC assessment in the same records says Exploitation: none for 101 of the 103.
  • OpenSSL publishes no CVSS score at all, and on NVD CISA’s is the only score for 75 of the 103, so any tool or policy that needs a number gets CISA’s. In 2026 alone, a “patch High and above in 30 days” rule fires 42 times on OpenSSL. OpenSSL rated 3 of those bugs High.

What CVE-2026-121481 actually needs

EVP_EncodeBlock() base64-encodes a buffer the caller hands it, into an output buffer the caller allocated. The caller is expected to reserve 4*((len+2)/3)+1 bytes. When the input length is close to INT_MAX, that arithmetic overflows a signed integer and the function writes past the end of the output buffer. That is a real defect, and the fix (public pull request #33652, opened on 10 November) is a sensible one.

The precondition is the whole story, though. To reach the overflow, an application has to pass EVP_EncodeBlock() an input of roughly two gigabytes. OpenSSL’s advisory text, which is the description NVD shows, is unusually direct about it:

  • OpenSSL’s own TLS, QUIC and CMS code never encodes buffers anywhere near that size.
  • The command line tools cannot practically be made to, because of the limits on reading a proxy password from a file and on bignum sizes.
  • No path by which an attacker can trigger it has been identified. An application is only exposed if it accepts an attacker-controlled length and forwards it unchecked.

The affected versions in the record are also narrower than a 9.8 implies: OpenSSL 4.0.0 to 4.0.3 and 4.1.0, with fixes due in 4.0.4 and 4.1.1. The 3.x lines that most production Linux servers still run are not listed as affected at all. Our own build box, a stock Ubuntu 24.04 LTS, reports OpenSSL 3.0.13.

Under OpenSSL’s published security policy, Low means issues that “require multiple unlikely conditions to be fulfilled”, which are fixed in public and shipped in the next scheduled release rather than an emergency one. That is exactly how this one is being handled.

Where the 9.8 came from

A CVE record is no longer one opinion. It is a set of containers: the CNA (here, OpenSSL itself) writes the description, the affected versions and its own severity, and Authorised Data Publishers can add enrichment alongside. CISA’s ADP, branded Vulnrichment, adds a CVSS score, an SSVC decision record and sometimes a CWE.

For CVE-2026-121481, the record shows OpenSSL publishing on 18 November with severity Low, and CISA’s ADP container updated on 19 November with:

CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H, base score 9.8.

Network-reachable, low complexity, no privileges, no user interaction, total compromise. That is the worst case of a heap overflow in a library function, scored with no knowledge of how any application calls it. In the very same container, CISA’s SSVC entry records Exploitation: none. The decision-relevant field and the headline number disagree, and only one of them fits in a news title.

On NVD, the record is still “Awaiting Analysis”, and the only CVSS metric displayed is CISA’s, labelled Secondary. There is no Primary score to sit next to it. That is why Hacker News, and every feed that scrapes NVD, says 9.8.

We checked 103 OpenSSL CVEs

OpenSSL’s vulnerabilities page lists every CVE the project has issued. We took every CVE ID from 2023 onwards, 102 records plus today’s (which the page does not list yet), pulled each one from the CVE Program’s public cvelistV5 repository, and compared the CNA severity with the CVSS 3.1 base score in CISA’s ADP container. Every one of the 103 had a CISA score.

OpenSSL severity CVEs CISA: Low CISA: Medium CISA: High CISA: Critical
Low 78 6 37 29 6
Moderate 20 0 4 14 2
High 5 0 0 5 0
Critical 0 0 0 0 0

Read the table by its empty corner. There is not one CVE where CISA’s band came out lower than OpenSSL’s. Eighty-eight came out higher. The eight CVEs CISA scored Critical were all rated Low or Moderate by the people who wrote the code, and OpenSSL has not rated any vulnerability Critical since CVE-2016-6309, ten years ago.

The pattern behind it is mechanical. Of the 78 Low-rated bugs, 43 were scored AV:N/AC:L/PR:N/UI:N: reachable from the network with no conditions at all. Nineteen of them are the same shape exactly, a denial of service scored A:H for a flat 7.5 High. OpenSSL’s policy, by contrast, explicitly puts “denial of service for less common protocols or on client side only” into Low.

Then the field that should calm everyone down: across all 103 records, CISA’s own SSVC entry says Exploitation: none for 101 and poc for 2. Nobody is exploiting these. The scores say otherwise because scores are not built to know.

Both of them are doing their job

It would be easy to call this score inflation and blame CISA. That would be unfair. A CVSS base score is defined as the intrinsic severity of a flaw, independent of any particular deployment, and for a library that means assuming the most exposed plausible caller. Given only “heap overflow in a widely used library”, 9.8 is a defensible base score. FIRST, which maintains CVSS, says in its own guidance that CVSS measures severity, not risk, and that the base score should be adjusted with environmental context.

OpenSSL is answering a different question. Its policy says plainly that CVSS “does not properly reflect the broad usage of the OpenSSL Library and does not fully account for likelihood of a configuration being affected”, so it does not use CVSS, and warns that scores from independent parties “might severely differ”. Its severity folds in how the library is actually called, what is on by default and how likely the preconditions are. That is closer to what a team deciding whether to page someone needs to know.

The failure is not in either rating. It is in pipelines that collapse two different questions into one number, and then take whichever number exists.

Why the number wins by default

Here is the structural problem. OpenSSL publishes a word, not a score: none of the 103 CNA records contains a CVSS metric. Scanners, dashboards and compliance reports want a number to sort, threshold and graph. When the vendor provides none, the only CVSS score in the record is CISA’s, so that is the one they use. NVD does not close the gap either: we queried all 103 on NVD’s API, and for 75 of them, including 66 that NVD marks as already analysed or modified, CISA’s is the only CVSS score shown. NVD published a Primary score of its own for just 26.

Now apply a perfectly normal policy. “Remediate High and Critical within 30 days; Critical within 7; fail the build at 9.0 and above.” Across the 72 OpenSSL CVEs published so far in 2026, CISA’s scores trigger the 30-day rule 42 times and the build-breaker 5 times. OpenSSL rated 3 of those bugs High and none Critical. Every one of the other 39 is an interrupt your team takes because a label was missing, and a team that is interrupted 42 times a year by OpenSSL alone stops reading the alerts, which is the outcome that actually hurts you the day a real Critical lands.

The volume matters too. 72 OpenSSL CVEs in under eleven months, 58 of them Low, against 31 published across the whole of 2023 to 2025 in our sample. Whatever is driving that increase, it means score-only triage costs more every quarter. We made a similar point from the other direction when we timed 6,119 advisories and found many landed long after the fix: the metadata around a vulnerability shapes the response as much as the vulnerability does.

What to change in your triage

  1. Gate on the vendor’s severity when the vendor publishes one. For OpenSSL, the Linux kernel and a growing list of mature projects, the CNA’s rating encodes knowledge of how the code is used. Keep the ADP score as a recorded field, not the trigger.
  2. Read the version range before the score. CVE-2026-121481 does not touch OpenSSL 3.x. If your fleet is on 3.0 or 3.5, the correct action is to note it and move on, whatever the number says. Check what your images actually link against, not what the base image claims; we covered why that distinction matters in our Brotli triage piece.
  3. Use exploitation status for urgency. SSVC Exploitation and CISA’s Known Exploited Vulnerabilities catalogue answer “is anyone doing this?” A 9.8 with Exploitation: none and a vendor Low is a scheduled-release fix, not an incident.
  4. Check the precondition in your own code. For this bug the question is small and answerable: does any of your C or native code call EVP_EncodeBlock() directly with a length an outside party controls? A grep across your own repositories answers it in minutes.
  5. Write the override down. If your policy says “fix 9.0+ in 7 days” and you decide not to, record why: vendor severity, affected range, exploitation status. Under NIS2-style supervision, a documented, reasoned decision is worth far more than a missed SLA with no explanation, and it stops the next engineer re-litigating the same CVE.

Where REPTILEHAUS fits

Most vulnerability programmes we are asked to look at are not short on data. They are short on judgement wired into the pipeline: which source to trust for which product, what the precondition looks like in this codebase, and when a red number is a scheduled patch rather than a weekend. We build and run production systems for clients across web, SaaS, AI and DevOps, and tuning dependency and vulnerability triage so it pages people for the right reasons is part of that work.

If your dashboard turned red today over an OpenSSL bug your servers do not even run, get in touch. We would rather help you set the rule once than watch your team chase 42 alerts to fix 3 bugs.

Methodology: CVE list from openssl-library.org/news/vulnerabilities, every CVE ID from 2023 onwards plus CVE-2026-121481; records from the CVEProject/cvelistV5 repository on 20 November 2026; CNA severity from the CNA container’s metrics.other field; CISA scores from the CISA-ADP container’s cvssV3_1 metric; bands per CVSS v3.1 (Low below 4.0, Medium 4.0 to 6.9, High 7.0 to 8.9, Critical 9.0 and above).

📷 Photo by Christian Lendl on Unsplash