81 Firefox Bugs, According to OpenAI: I Counted What Mozilla Has Published
On November 5, OpenAI introduced GPT-6 Cyber. The announcement says its researchers used the model to identify 81 previously unknown Firefox vulnerabilities, all high severity according to Mozilla.
That is a figure from the company that sells the model, and the announcement links a single Mozilla advisory as the public record. I wanted to know how much of that 81 can be seen from the outside. On November 16 I downloaded all 129 security advisories Mozilla has published in 2026 and started counting.
The short result: 89 CVEs carry a credit for a reporter affiliated with OpenAI, and 85 of them have high impact. That sounds like a match for the 81. It is not a confirmation, and the reasons are the most useful part of this analysis.
What exactly does OpenAI claim?
According to the announcement, during a research campaign GPT-6 Cyber supplied evidence and reproducible tests for 81 new vulnerabilities that Mozilla classified as high severity. OpenAI adds that another 30 already known issues turned out to be more severe than previously reported.
It also says it handed the findings to Mozilla with proposed patches and that Mozilla began shipping fixes soon after. The text speaks of fixes for many of the vulnerabilities, not all of them, and links advisory MFSA 2026-118 as the reference.
There are three things the announcement does not include. It does not list the identifiers of the 81, it does not say when the campaign started, and it does not explain whether it counts bugs or CVEs.
The system card that OpenAI links does not clarify this either. It is a document about evaluations and safeguards, and in the text the page loads I found no mention of Firefox or Mozilla.
The announcement also includes a testimonial from Brian Grinstead, Distinguished Engineer for Firefox at Mozilla, saying the model found previously unknown serious vulnerabilities and helped ship fixes. It is hosted on OpenAI’s site and contains no figure. Help Net Security attributes the numbers to OpenAI and does not report an independent verification of the total.
How did I count, and what does each number mean?
Each Mozilla advisory lists CVEs with two fields I can use: the impact Mozilla assigns and the reporter (the field is called Reporter) credited as the discoverer. I downloaded advisories MFSA 2026-01 through MFSA 2026-129, which are all the 2026 advisories shown on Mozilla’s advisory index.
The same CVE appears in the Firefox, Firefox ESR and Thunderbird advisories, so I count each CVE once and give it the date of the first advisory that lists it. I count it as an OpenAI credit if its reporter field contains the word OpenAI.
This is the exact definition I use throughout the article: unique CVEs, de-duplicated across the Firefox, ESR and Thunderbird advisories, whose Reporter field names an OpenAI-affiliated reporter, from MFSA 2026-01 to MFSA 2026-129, downloaded on November 16, 2026. In total, those 129 advisories add up to 969 unique CVEs.
I ran three checks. The word OpenAI appears 377 times in the HTML of the 129 advisories, always inside a reporter field. There is no other spelling, no Open AI, no GPT, no Daybreak.
The third: each of the 89 CVEs links to a single Bugzilla bug. In all of 2026 only 1 of the 969 CVEs groups two bugs, so counting CVEs or counting bugs gives the same result.
There is one limitation that conditions everything else: the reporter field names a person, a team or an alias, never a model. Four different strings appear: OpenAI Codex Security (kylebot) on 59 CVEs, Amy Burnett of OpenAI on 28, and two single credits, one from OpenAI Preparedness and one from Trail of Bits in partnership with OpenAI.
A credit does not tell me whether the finding came from GPT-6 Cyber, from an earlier model or from a person with other tools.
What do Mozilla’s advisories show?
The table groups the 89 CVEs by the date of the first advisory that lists them. The Other column is 3 low-impact CVEs and 1 moderate-impact CVE.
| First advisory | Firefox (advisory) | High impact | Other | Total | Cumulative high |
|---|---|---|---|---|---|
| May 12 | 150.0.3 (MFSA 2026-45) | 1 | 0 | 1 | 1 |
| July 21 | 153 (MFSA 2026-68) | 2 | 2 | 4 | 3 |
| August 18 | 154 (MFSA 2026-74) | 4 | 1 | 5 | 7 |
| September 15 | 156 (MFSA 2026-90) | 0 | 1 | 1 | 7 |
| September 29 | 157 (MFSA 2026-97) | 2 | 0 | 2 | 9 |
| October 13 | 158 (MFSA 2026-106) | 9 | 0 | 9 | 18 |
| October 27 | 159 (MFSA 2026-118) | 41 | 0 | 41 | 59 |
| November 10 | 160 (MFSA 2026-125) | 26 | 0 | 26 | 85 |
| Total | 85 | 4 | 89 |
Source: Mozilla security advisories, my own count from data downloaded on November 16, 2026. It is my count of public data, not a figure from Mozilla or OpenAI.
The advisory OpenAI links, MFSA 2026-118, covers Firefox 159 (October 27). It lists 167 CVEs and 41 carry an OpenAI credit, all high impact.
On November 5, when OpenAI published the announcement, Mozilla’s latest advisory was the one for Firefox 159.0.1 (November 3) and it carried no OpenAI credits. Up to that date the total was 63 CVEs with an OpenAI credit, 59 of them high impact.
Five days later, MFSA 2026-125 (Firefox 160, November 10) listed 99 CVEs and 26 carry an OpenAI credit, all high. Firefox 160 shipped that day, according to Mozilla’s release data. With that advisory the total reached 89, with 85 high.
The 22 credits before October 27 always name specific people or a specific team: 20 are Amy Burnett of OpenAI and two are single credits. The OpenAI Codex Security (kylebot) label first appears on October 27 and accounts for 59 of the 67 credits in the last two releases. OpenAI does not explain that change of label and I am not going to guess.
Why doesn’t 85 confirm 81, but doesn’t contradict it either?
Because a count is not a list. If the 81 were a natural cut of the data, I would see it.
Counting only high-impact CVEs gives 85 in total, 82 since August 18, 76 since October 13, 67 since October 27 and 59 with the OpenAI Codex Security (kylebot) label. No cut gives 81.
There are several reasons why the 81 and what shows up in the advisories can differ, and from the outside I cannot tell them apart:
- Unknown dates. OpenAI does not say when the campaign started. Credits dated before the announcement may or may not belong to it.
- Findings credited to someone else. A bug Mozilla fixed under another name in the reporter field does not enter my count. One still under embargo does not either, because it has no advisory yet.
- Credits unrelated to the 81. Part of the 85 may come from earlier work, or from work other than GPT-6 Cyber.
The difference between bugs and CVEs is not the explanation, as I checked earlier: in 2026 only 1 of every 969 CVEs groups more than one bug. Nor can I check the 30 issues OpenAI says were more severe than reported, because an advisory does not record that a severity was revised.
Here is what I can say. Mozilla has published many fixes credited to OpenAI: 63 CVEs when the announcement came out and 89 today. That fits the announcement’s statement that fixes for many of the vulnerabilities were already out.
What I cannot say is that those 85 are the 81, or that GPT-6 Cyber found any specific one of them.
What would settle it is the list of identifiers of the 81, published by OpenAI or by Mozilla. With it, checking would take an afternoon. In the meantime I will keep watching the advisories of upcoming releases.
What about the 948 Linux exploits?
On November 4 OpenAI also published A thousand fewer paths to root, about an autonomous pipeline that finds, validates and patches vulnerabilities in the Linux kernel. What follows are OpenAI’s figures, measured by OpenAI and not independently replicated.
| Stage | OpenAI figure |
|---|---|
| Finding candidates after deduplication | 2,308 |
| Vulnerabilities reproduced in virtual machines | 1,724 (74.7%) |
| Local escalation candidates that moved to exploitation analysis | 1,238 |
| Local privilege escalation exploits that worked | 1,169 |
| Distinct vulnerabilities behind those exploits | 948 |
| Critical findings reported to maintainers | 304 |
Source: OpenAI’s article of November 4 and its announcement of November 5. The table leaves out the information-leak branch (313 candidates and 291 proofs of concept).
I checked what can be checked without access to their data. The five per-stage costs it details add up to $1,213,428, against the $1,213,430 it states in total, and 1,238 candidates minus 69 failures gives the 1,169 exploits. That shows the arithmetic is consistent, not that the results are true.
The limits the article itself lets you see:
- It is local escalation. An exploit of this kind already needs an unprivileged account on the machine. It is not remote code execution.
- Lab environment. The tests ran in virtual machines with a stock desktop and the latest kernel from the main branch in October, with the usual mitigations enabled.
- Not everything was still open. When the ten days of analysis ended, according to OpenAI, 61 of those escalations were already patched in the kernel, 43 had patches in staging and 299 were being discussed in public. That is why it recommends revalidating each finding against the latest code.
- A success was not always the assigned success. In an earlier run, 20 of 200 exploits counted as good had reached root through a different vulnerability than the assigned one. A human review caught it and OpenAI added an independent check.
- The model’s name is missing from the article. The November 4 text speaks of an upcoming cyber model and does not name GPT-6 Cyber. The November 5 announcement does attribute the 948 to GPT-6 Cyber and links that article, so it is consistent, but the attribution appears only in the announcement.
- There is no public list. Unlike Firefox, I have not found any list of the 948 that would let you count them from the outside. OpenAI says it has reported 304 critical findings to maintainers, only part of the 948.
OpenAI also calculates a hypothetical value of between $18.9 million and $66.3 million by applying historical hacking-contest rewards to the 948 exploits, roughly 16 to 55 times the cost. It is an exercise with contest prices, not money collected, and I would not rely on it.
The conclusion that convinces me most comes from its own figures. Finding candidates cost 26% of the total, about $320,000, and validating, assessing and exploiting them the remaining 74%: the bottleneck was not discovering, it was checking.
Its practical recommendation, which I share, is that an unprivileged account on an unpatched system is equivalent to root for a well-resourced attacker.
What do I take away from this?
Four things, in order of usefulness:
- Found by AI is a provenance label, not a severity or an auditable figure. Always ask for the unit of count, the date window and the list of identifiers.
- Advisories are a primary source you can count. A reporter field and a short script are enough to see how much of a company’s work shows up in a product.
- Update. Firefox 160 fixes 26 high-impact CVEs credited to OpenAI, and the Linux kernel is being patched at a pace that no longer forgives lagging systems.
- Do not compare figures of a different nature. A campaign’s 81 and a public 85 of credits can be the same set, overlapping sets or mostly different sets, and the numbers alone do not tell you which.
The counting script (Python, standard library only) and the CSV with the 89 CVEs are in the editorial assets folder of this blog’s repository.
I wrote about another case of a kernel exploit found by a model in The Linux Kernel Race Behind Anthropic’s Root Exploit.