Google's Chrome 157 release fixes 213 security bugs, and 130 of them — 61.0% — were reported by Google itself. The AI-lab share everyone is excited about is 21 of 213, or 9.9%.

I did not take that from a news summary. I fetched Google's official Chrome Releases post for the November 3 stable update (157.0.8103.38 for Linux, .38/.39 for Windows and Mac), parsed every entry of the form "[bounty][bug id] Severity CVE-…: flaw in component. Reported by X on date", and deduplicated by CVE number. The deduplicated count lands exactly on Google's stated total, which is the first check that the parse is sane.

parsed entries (raw, page repeats the list): 426 unique CVEs after dedupe: 213 Google's stated total: 213 ✓ severity breakdown (unique CVEs): Critical: 7 ( 3.3%) High: 77 (36.2%) Medium: 75 (35.2%) Low: 54 (25.4%) total: 213

The severity split matches the tallies in the secondary coverage, which is a second, independent check. The interesting part is who found them.

reported by (unique CVEs, top groups): Google (internal): 130 (61.0%) Xinyang Ge, Anthropic, assisted by Claude: 11 Xinyang Ge (no assistance noted): 9 xinyang (lowercase credit): 2 Vlad Stolyarov, Anthropic, assisted by Claude: 3 OpenAI Codex Security: 7 everyone else combined: 51 Claude-assisted, explicitly credited: 14 ( 6.6%) Codex Security, explicitly credited: 7 ( 3.3%) AI-credited total: 21 ( 9.9%)

Two counting notes, because this is exactly where a lazy summary goes wrong. First, "Xinyang Ge" appears under three credit strings, so I counted 22 reports under that name only by adding 11 + 9 + 2 — and the lowercase "xinyang" entries might or might not be the same person, so treat 22 as a name-string total, not a proven person total. Second, "assisted by Claude" is an explicit credit in the post, not my inference: 14 entries carry it. The Codex entries carry no "assisted by" wording; they are credited to "OpenAI Codex Security" with a handle in parentheses, so 7 is a credit count, not a claim about how much of the work the tool did.

The seven Criticals tell a smaller version of the same story:

CVE-2026-121156 Buffer overflow in Dawn Zhenpeng (Leo) Lin, depthfirst, 2026-09-30 ($10,000) CVE-2026-121229 Memory corruption in ANGLE @mfx, 2026-09-17 ($1,000) CVE-2026-121246 Use after free in WebGL Google, 2026-09-08 CVE-2026-121035 Use after free in Prerender Xinyang Ge (Anthropic), assisted by Claude, 2026-09-28 CVE-2026-121247 Buffer overflow in Dawn Google, 2026-10-03 CVE-2026-121203 Use after free in Geometry Xinyang Ge, 2026-10-07 CVE-2026-121126 Use after free in Core Wenbin Liu, 2026-10-08

Four of the seven are use-after-free, and four of the seven sit in the GPU stack (Dawn twice, ANGLE, WebGL). Across all 213, the most common flaw class is also use-after-free at 45 reports (21.1%), followed by incorrect authorization at 37 (17.4%). That is a memory-safety release wearing an authorization-release trench coat: the headline count is dominated by one old bug class, not by some exotic new attack surface.

What the post does not say matters too. The segment I parsed contains no claim that any of these bugs is exploited in the wild, and Google's own note says bug details and links "may be kept restricted until a majority of users are updated with a fix." So the honest reading is: 213 fixes, most found internally, a real but minority AI-assisted slice, and no public evidence yet about exploitation. Anyone telling you this release proves AI has taken over vulnerability research is reading 21 as if it were 213.

The one number I would actually watch next release is not the AI share. It is whether Google's internal 61% holds. If internal reports fall while the total stays above 200, that means the outside world — humans, AI-assisted or not — is finding a growing share of Chrome's bugs. This release, the internal team still found nearly two out of three.

Sources: Google Chrome Releases, "Stable Channel Update for Desktop," November 3, 2026 — https://chromereleases.googleblog.com/2026/11/stable-channel-update-for-desktop.html ; all counts above are from my own parse of that post on November 4, 2026.