Cybersecurity · AltTab Dad

Anthropic's unreleased Claude wrote working exploit chains for Chrome, Safari, Android, Windows and Linux — update Chrome to 156

Quotable summary

On 27 October 2026 Anthropic’s Frontier Red Team published five exploit chains written by an unreleased, internal Claude research model, each turning newly found bugs into real attacks with little or no user interaction: one-click code execution outside Chrome’s sandbox, a one-click Safari same-origin bypass, an Android lock-screen bypass that pulls photos off a locked phone, a remote crash of Windows HTTP/3 servers (CVE-2026-33096), and zero-click root on Linux servers with a multipath TCP listener (CVE-2026-98432). All were reported under coordinated disclosure and Anthropic says each is fixed. Google’s Chrome 156 stable release the same day carries 52 security fixes, several credited to “Xinyang Ge (Anthropic), assisted by Claude.” Action: update Chrome to 156.0.8078.73 or later and restart the browser.

What changed

Target What Claude built Status per Anthropic
Chrome (macOS, default config) One-click RCE outside the sandbox: V8 heap corruption, V8 sandbox escape + ASLR bypass, CacheStorage race, FedCM use-after-free Patched as of Chrome 156
Safari One-click same-origin-policy bypass reading history, cookies and session data; data-only attack via JavaScriptCore use-after-free and a sandbox IPC flaw Fixed by Apple
Android (tested on Pixel 9) Lock-screen bypass over USB extracting photos and personal data without unlocking Addressed by Google
Windows Remote crash of any machine serving native Windows HTTP/3 via a QUIC kernel out-of-bounds read, CVE-2026-33096 Resolved by Microsoft
Linux kernel Zero-click remote root via multipath TCP race/use-after-free, CVE-2026-98432 Patched upstream and backported to maintained stable kernels

Chrome 156 numbers

Item Figure Source
Stable release 156.0.8078.73/.74 Windows and Mac, 156.0.8078.73 Linux, 27 Oct 2026 Chrome Releases
Security fixes in this release 52 Chrome Releases
CVEs from Anthropic’s findings 62, all critical or high, incl. 13 sandbox escapes and 21 renderer compromises, patched “as of Chrome 156” Anthropic

The two counts measure different things: Google’s 52 is one release’s fix list, while Anthropic’s 62 covers all its reported Chrome CVEs fixed up to and including 156. Not every fix in Google’s list credits Claude.

Why it matters

FAQ

Is my browser safe? Update Chrome to 156.0.8078.73 (Linux) or 156.0.8078.73/.74 (Windows and Mac) or later, and keep Safari current. Anthropic says Apple’s fix is already available.

Were these exploited in the wild? Neither Anthropic’s post nor Google’s release notes report in-the-wild exploitation of these bugs. Anthropic reported them under coordinated disclosure.

Which Claude model did this? Anthropic calls it an “unreleased, internal research model.” It has not named it publicly.

Can defenders get these capabilities? Anthropic points qualifying security professionals to its expanded Cyber Verification Program and recommends scanning code at pull-request time.

Sources

Source Role Link
Anthropic Primary: five exploit chains, CVE counts, disclosure status Anthropic — Discovering and exploiting vulnerabilities with Claude
Google Chrome Releases Primary: Chrome 156 versions, 52 fixes, credits Chrome Releases — October 2026
The Hacker News Secondary: coverage of the chains The Hacker News — Anthropic’s unreleased model chains flaws
Cybersecurity News Secondary: Chrome chain walkthrough Cybersecurity News — Claude builds Chrome and Linux exploit chains