Chrome's 27 October update credited Anthropic or OpenAI's Codex Security on 12 of 52 fixes. The credits cover AI-related reporting only and give no overall share of AI in bug discovery.
Chrome's 27 October stable update lists 52 security fixes. Twelve credit lines name Anthropic or OpenAI's Codex Security, which is 23 percent of the release.
What the credit lines show
Six lines say "assisted by Claude", two name Anthropic with Ada Logics, and four name OpenAI Codex Security. Ten more name the Anthropic researcher Xinyang Ge alone, with no tool or employer stated. The count of 12 leaves those ten out, so it counts attributions and not every AI-assisted find.
The share swings from release to release. Chrome's 20 October update, which promoted Chrome 156, carried 353 fixes, and 9 credit lines name Anthropic, Claude or Codex Security (2.5 percent). The 29 October update had 21 fixes and none.
Across those three updates, AI-related names appear on 21 of 426 fixes. That is 4.9 percent, on CTOL's arithmetic. The 27 October share is far above it.
The 29 October update's one Critical fix, in MediaStream, was reported by Xinyang Ge on 21 October, eight days before release. The credit line carries no Anthropic or Claude label.
A single release shows a batch and cannot show a trend. Two of the 27 October batch's four Critical fixes, CVE-2026-118306 and CVE-2026-118331, hold the "assisted by Claude" credit.
How long the fixes took
A second measure is the interval between the report to Google and the stable release that contains the fix. It includes investigation, fixing, testing and scheduling, so it does not measure time spent unfixed. Each credit line gives the report date.
For the 12 named fixes in the 27 October release, the interval ran from 12 to 29 days, with a median of 16.5. The two Critical fixes took 20 and 16 days. The slowest was an OpenAI Codex Security report, CVE-2026-118350, at 29 days.
The other 40 fixes in the same release ran from 6 to 91 days, with a median of 12.5. The two groups overlap, and their medians differ by four days. The figures do not show how Google triages AI-credited reports compared with others.
What the earlier AI exploit research covers
Anthropic's June study looked at exploiting publicly disclosed fixes. It did not involve Chrome or the search for unknown bugs.
Anthropic's study of 8 June tested "patch diffing", where an attacker compares software before and after a fix to find the bug it closed. Its Mythos Preview model built 8 working exploits from 18 Firefox patches, the first in just under an hour. On 21 Windows kernel patches it built 8 full attack chains that reached SYSTEM control, at about $2,000 each. Anthropic concludes that "N-hour is closer to the reality" than "N-day".
Those are the producer's own results. The Firefox runs used the JavaScript engine alone in a test shell, not the full browser. Microsoft had rated 14 of the 21 Windows bugs unlikely to be exploited, and Mythos made proofs of concept for 13, including a privilege escalation for one. A proof of concept can show anything from a crash to privilege escalation without being a reliable, end-to-end attack. No record read shows an AI agent turning a Chrome patch into an exploit.
The Firefox clock also differs from Chrome's. Mythos produced a first exploit within an hour of Mozilla's patch, while the patched Firefox 148 was 18 days from release. Across the studied patches the median gap was 19 days, and Mozilla has announced weekly scheduled dot-release slots starting with Firefox 151.
Chrome's clock starts when the bug is reported, so the two cannot be stacked into one timeline. The 18 days belongs to a Firefox lab case and to no Chrome bug. Chrome's own note says access to bug details "may be kept restricted until a majority of users are updated with a fix".
October's Chrome news is about finding unknown bugs and building attacks on them. XBOW says it built a full Chrome exploit chain in six days with a public OpenAI model and won a $250,000 Google reward. It was the first reward under rules Google introduced in April 2026.
Anthropic reports that an unreleased model has found Chrome bugs now carrying 62 CVEs, all patched as of Chrome 156. Project Zero counted 45 actionable V8 bugs from Google's Big Sleep agent against 37 from its Fuzzilli fuzzer over six months. In production only a quarter of the bugs overlapped, so each method finds bugs the other misses. Project Zero recommends that vendors "perform both fuzzing and agentic analysis of their code".
Limits and risks
Anthropic names its own limits. Exploit development is one step in an attack, and delivery, finding targets and evading defences remain. XBOW's timeline is self-reported and humans steered the strategy. It still believes agents alone could find each of its three bugs within a year.
XBOW identifies a separate risk of duplicate discovery. The first version of its chain turned out to duplicate a bug Google already knew. The authors add that "bug collisions are a major risk" now that everyone is pointing language models at codebases like Chrome. Chrome's credit lines do not show how often such collisions happen.
XBOW reports $40,000 in compute for a chain that won a $250,000 reward, which is 16 percent of the payout. That excludes human labour and other costs, and it is not a measure of profit. XBOW says a team of four needed 6 to 9 months for such a chain before language models.
What the next data can show
CTOL's view is that 12 of 52 on 27 October is a high concentration. It sits against 21 of 426 across the three October updates examined. The credit lines do not establish AI's overall share of vulnerability discovery.
The 29 October update already shows how sharply the share can swing between batches. Later stable releases can show whether similarly concentrated credits recur, but several would be needed before inferring a lasting change. The Chrome 157 stable date is not confirmed, and a missing credit line does not show a bug was found by hand.