On October 27, Google released a Chrome security update fixing 52 flaws, eighteen of which are credited to Xinyang Ge, an Anthropic researcher, or to Anthropic, meaning more than one in three. The same day, Anthropic’s Frontier Red Team published a blog post to which Xinyang Ge is one of seven contributors.
This post presents an internal, non-commercial research model as the author of full attack chains against Chrome, Safari, Android, and Linux, as well as a denial-of-service flaw on Windows. Anthropic writes that “62 CVEs have been assigned and fixed” in Chrome up to version 156, all of critical or high severity. A CVE is a public identifier assigned to a security vulnerability.
ActuIA reviewed the eleven Chrome security bulletins for computers published by Google between September 15 and October 29, 2026. Of the 929 listed flaws, 77 bear the name of Xinyang Ge or Anthropic. The first report dates from September 10, the last from October 23, and none of these flaws show a bounty to date.
A variable share depending on the bulletins
In October, Google published seven Chrome security updates for computers, on the 1st, 6th, 13th, 15th, 20th, 27th, and 29th. They list 731 flaws, 72 of which are credited to Anthropic or its researcher, or about one in ten. The proportion varies significantly from one bulletin to another. It remains low in the two version updates, with 14 flaws out of 247 for Chrome 155 on October 6 and 15 out of 353 for Chrome 156 on October 20. In several intermediate patches, it exceeds one third, such as on October 13 (15 flaws out of 37) or the 27th (18 out of 52).
Except for two lines, the 77 flaws cite the same researcher. Forty-nine are credited to “Xinyang Ge” with no other mention, 26 to “Xinyang Ge (Anthropic), assisted by Claude” and two to “Anthropic and Ada Logics”, a security consulting firm. Eighteen are classified as critical and 59 as high. In the reward column, all show “N/A” or “TBD”, meaning to be determined.
ActuIA’s count shows 55 flaws credited by the release of Chrome 156 on October 20, and 73 on October 27, the day of the blog post. The 62 announced by Anthropic fall between the two. The company does not publish the list of these CVEs, and Google only mentions Claude on 26 lines, so a line-by-line comparison is not possible. Google also indicates that access to details of some flaws may remain restricted “until a majority of users have received the fix”.
Patches outside of Chrome
Outside of Chrome, the references given by Anthropic point to patches published before the blog post. The Windows flaw, CVE-2026-33096, was fixed during the Patch Tuesday of April 14, 2026. It affects the HTTP.sys driver and allows remote denial of service without authentication, with a CVSS score of 7.5, according to Tenable’s sheet. Anthropic estimates “several million” exposed servers. The Linux kernel flaw, CVE-2026-98432, was published on October 15 with a score of 9.8 and fixed in all maintained stable branches, according to Anthropic.
The Apple bulletin cited by Anthropic concerns iOS and iPadOS 26.5.2, released on June 29. It credits a WebKit flaw to “Milad Nasr and Nicholas Carlini with Claude, Anthropic”, two other contributors to the blog post, as well as several flaws to OpenAI Codex Security. For Android, Anthropic writes only that the issue was “addressed by Google”, with no CVE number.
Twelve days median
For the 77 Chrome flaws, ActuIA calculated the gap between the report date and the bulletin publication: twelve days median, with a minimum of 3 days and a maximum of 41. In October, two security updates are never more than seven days apart. Anthropic gives vendors 90 days to fix, then waits 45 more days before publishing details that go beyond the security advisory. Its blog post recommends developers analyze code on every merge request, and judges that analyzing one’s own code is “more important than ever”. The version required as of October 29 is 156.0.8078.79 on Linux, and 156.0.8078.79 or 156.0.8078.80 on Windows and Mac.
