sysnote
Analysis

We recounted Mozilla's advisories: 78 high-severity Firefox CVEs credited to OpenAI, against GPT-6 Cyber's claimed 81

OpenAI's claim that GPT-6 Cyber found 81 previously unknown high-severity Firefox vulnerabilities can now be checked against Mozilla's own security advisories, and it mostly holds up. Firefox 159 (MFSA 2026-118, October 27) and Firefox 160 (MFSA 2026-125, November 10) credit OpenAI reporters with 67 CVEs, every one rated high. Counting the same reporters' bugs back to Firefox 157, the total is 78, three short of OpenAI's number.

When we covered the GPT-6 Cyber launch on November 5, the Firefox figure was OpenAI's word plus a quote from Firefox engineer Brian Grinstead. Mozilla has still not published its own tally. But its advisories list a reporter and an impact level for every CVE, so we downloaded every 2026 advisory and counted.

The count

Firefox release Advisory Date OpenAI-credited CVEs All high-impact CVEs in release
Firefox 157 MFSA 2026-97 Sep 29 2 (Amy Burnett) 38
Firefox 158 MFSA 2026-106 Oct 13 9 (Amy Burnett) 30
Firefox 159 MFSA 2026-118 Oct 27 41 (34 Codex Security, 7 Amy Burnett) 108
Firefox 160 MFSA 2026-125 Nov 10 26 (25 Codex Security, 1 Amy Burnett) 57

All 78 are rated high. None is critical, moderate or low, and none of these advisories rates anything critical. Two credit strings appear: "OpenAI Codex Security (kylebot)" and "Amy Burnett of OpenAI". OpenAI-credited bugs make up 38% of the high-impact CVEs in Firefox 159 and 46% in Firefox 160. Two examples: CVE-2026-117921, a use-after-free in CSS parsing fixed in Firefox 159, and CVE-2026-124872, a use-after-free in the cycle collector fixed in Firefox 160.

The bugs are mostly memory-safety work. Use-after-free accounts for 45 of the 59 Codex Security CVEs. Amy Burnett's 19 lean toward the JavaScript engine, including six JIT miscompilations.

Why 78 and not 67

Mozilla's credits name people and products, not models. Nothing in an advisory says "GPT-6 Cyber", so tying a CVE to the model is our inference. The evidence for the Firefox 157 and 158 bugs is the bug numbers. Amy Burnett's two CVEs in Firefox 157 (bugs 2073266 and 2073268) and nine in Firefox 158 (2073267–2076269) come from the same short run of Bugzilla IDs as her Firefox 159 and 160 bugs (2074167–2076271). They look like one batch of reports, shipped over four releases. The Codex Security bugs form a second, later batch (2079687–2086062), all fixed in Firefox 159 and 160. The bugs themselves are access-restricted, so we can't see their filing dates or text.

Older OpenAI credits look like different work. They go back to a single high CVE in Firefox 150.0.3 in May, credited to "OpenAI Preparedness, Bill Demirkapi". Firefox 153 and 154 credit Amy Burnett and "Artur Cygan of Trail of Bits in partnership with OpenAI", with a mix of high and low ratings. Across the whole year, Mozilla's advisories credit OpenAI-affiliated reporters with 89 unique CVEs: 85 high, one moderate and three low.

The "kylebot" in the Codex Security credit is not explained anywhere. It is one character off @ky1ebot, the OpenAI researcher Clint Gibler credited with building the pipeline behind the 948 Linux kernel exploits. That fits, but it is a resemblance, not a confirmation.

Don't count the ESR copies

Firefox ESR and Thunderbird advisories repeat the same CVE IDs. ESR 153.6 lists 33 OpenAI-credited CVEs and Thunderbird 159 lists 41, and the November ESR 153.7 and Thunderbird 160 advisories repeat 24 and 26 more. Added together across the October and November releases, they make well over 200 "OpenAI" entries. Every one is a copy of a CVE already in a mainline Firefox advisory. None is unique to ESR or Thunderbird.

One more correction to a common reading. Mozilla's advisories now carry a note that it issues a separate CVE for each internally found memory-safety bug instead of rolling them into one. Some coverage presented this as a response to the OpenAI wave. The note announcing the change first appears in the Firefox 156 advisories on September 15, before any of these batches shipped.

What can't be checked

The 30 "severity upgrades". OpenAI says GPT-6 Cyber showed that 30 known issues were more severe than previously reported. Advisories show only the final impact rating, with no history. An upgraded bug keeps its original reporter, and the Bugzilla entries where a re-rating would show up are closed to the public. This half of the claim can't be verified from outside.

Mozilla's own account. Mozilla has written about AI bug-finding, just not about this campaign. In an October 29 post, Bobby Holley says the Firefox team has used early access to frontier models "from Anthropic and OpenAI" since February. He says Firefox has shipped more than seven times as many security fixes as in the same period of 2025. The post doesn't name GPT-6 Cyber or break the fixes down by lab. Its real subject is the patch gap. Mozilla says it has seen evidence that GLM 5.3 can exploit some fixes for bugs found by stronger models, so it now keeps regression tests confidential until users have the update. Anthropic is in the same race: its Claude Mythos 5.2 launch came with a partner claim about Firefox vulnerability discovery.

What's left

Three of OpenAI's 81 are not in any advisory yet, or 14 if you count only Firefox 159 and 160. Firefox 161 is due on November 24 on Mozilla's release calendar, and its advisory is the next place to look.

The other promise from November 5 hasn't arrived either. OpenAI said GPT-6 Cyber would expand "soon" from a select group to all Daybreak Red customers. What it shipped instead at its November 12 "Cyber at the Frontier" keynote were two products, Daybreak Red Team and Daybreak Defense Factory, announced by Greg Brockman. We have not seen an announcement that GPT-6 Cyber itself has reached more customers.

For a vendor security claim, this one checks out unusually well. It names a codebase, a severity level and a number, and the vendor's own advisories bear out 78 of the 81 at the stated severity.

Comments

No comments yet. Be the first.

Write a comment

Comments are published after review.