NEWS
Rival AIs Found Chrome’s Critical Bugs Before Carbon
Chrome 156.0.8078.73 patches 52 bugs, four critical and all from one Anthropic researcher, the day before Google gated Gemini 4 Carbon to cyber defenders.
Google released Chrome 156.0.8078.73 on October 27 with 52 security fixes, four of them rated Critical. All four critical bugs were reported by researcher Xinyang Ge, and two of those credits name Anthropic and assistance from Claude.
On October 28, Google announced Gemini 4 Carbon, a frontier model it is first giving to a subset of cyber defenders. The patch note and the model launch arrived one day apart, and they do not tell the same story about who is finding Chrome’s worst bugs.
The October 27 Build That Closed 52 Chrome Flaws
The stable channel moved to 156.0.8078.73 and 156.0.8078.74 on Windows and Mac, and to 156.0.8078.73 on Linux, in a note signed by Srinivas Sista. Google said the build will roll out over the coming days and weeks. It did not say any of the bugs had been used in the wild, and it said it may hold back bug details until a majority of users are patched.
THE OCTOBER 27 DESKTOP TOTAL
- 52 fixes: Google listed 52 security fixes in this build, covering file handling, layout, graphics, JavaScript, media, and UI.
- Four critical: FileSystem memory corruption, FileAPI input validation, Skia input validation, and a Layout use-after-free.
- 46 high, two medium: The rest of the list is 46 High and 2 Medium, with V8, Skia, SVG, and FileSystem showing up more than once.
- Staged rollout: Desktop, Android, and managed fleets do not all move on the day the note is posted.
Chrome for Android 156.0.8078.73 carries the same security fixes unless Google says otherwise, and Google Play was set to offer it over the next few days after October 27. That is a different clock from desktop, and it is the one most phones actually wait on.
The October 27 note is also not Chrome 156’s first security drop. Chrome Releases listed 353 security fixes when 156 first hit the stable channel on October 21. This follow-up is the build that closed the four new critical bugs, and it is the one enterprises still have to confirm on disk.

Xinyang Ge Filed Every Critical Bug
Ge’s name is on all four critical entries. Two of them, CVE-2026-118306 in FileSystem (reported October 7) and CVE-2026-118331 in FileAPI (reported October 11), are credited as Xinyang Ge (Anthropic), assisted by Claude. The other two, CVE-2026-118348 in Skia and CVE-2026-118344 in Layout, were reported by Ge on October 17 with no Anthropic line on the credit.
THE FOUR CRITICAL ENTRIES
| CVE | Component | Bug type | Credit and date |
|---|---|---|---|
| CVE-2026-118306 | FileSystem | Memory corruption | Xinyang Ge (Anthropic), assisted by Claude, Oct 7 |
| CVE-2026-118331 | FileAPI | Improper input validation | Xinyang Ge (Anthropic), assisted by Claude, Oct 11 |
| CVE-2026-118348 | Skia | Improper input validation | Xinyang Ge, Oct 17 |
| CVE-2026-118344 | Layout | Use after free | Xinyang Ge, Oct 17 |
Across the full 52, Ge’s name appears 16 times. Six of those 16, including the two critical FileSystem and FileAPI bugs, carry the Anthropic and Claude tag. OpenAI’s Codex Security is on four High bugs (Turbolev, ImageCapture, Video, PDFium). Google itself is on 19, which is the largest single pile, and almost all of those are High or Medium, not Critical.
WHO ELSE IS ON THE 52
- Google: 19 bugs, none of the four criticals.
- Xinyang Ge: 16 bugs, including every critical, six with Claude assistance.
- OpenAI Codex Security: 4 High bugs, credited to handles amyb and kylebot.
- Anthropic and Ada Logics: 2 joint High reports that do not use Ge’s name.
- Other outside reporters: Arni Hardarson of Neonix Security (2), malichamdan (2), plus single credits for Jihyeon Jeong at Seoul National University’s Compsec Lab, Mozilla, and several named or anonymous filers.
Google marked a large share of the Ge, Codex, and internal items as N/A for a payout, and left independent reports as TBD. The credit lines are the part of the note that will last after the version string ages out, because they show which lab’s tools were already inside Chrome’s bug tracker in October.
The Proof of Concept Launched Calculator on macOS
The four critical bugs in the October 27 note are not a four-step exploit chain. They sit in FileSystem, FileAPI, Skia, and Layout. The chain Anthropic described the same day uses a different set of Chrome pieces, and it is the claim that makes the patch note look small.
Anthropic’s Frontier Red Team published a research post on October 27 describing work with an unreleased internal model. After initial prompting, the post says, Claude wrote a full exploit chain for one-click remote code execution in Chrome. The proof of concept, on a default Chrome install on macOS, launched the Calculator app within seconds of a page loading, with no further clicks.
THE FOUR STEPS ANTHROPIC DESCRIBED
- V8 heap: A bug in Chrome’s JavaScript engine that let a malicious page corrupt V8 heap memory.
- V8 sandbox: A second bug that broke out of V8’s memory region, defeated ASLR, and used a return-oriented programming chain to take over the renderer.
- CacheStorage race: A race that let the compromised renderer read memory outside Chrome’s process sandbox and map the privileged browser process.
- FedCM use-after-free: A use-after-free in Chrome’s federated sign-in feature that ran code outside the sandbox.
After initial prompting, Claude wrote a full exploit chain enabling an attacker to execute arbitrary code as the user, outside the browser’s sandbox.
Anthropic Frontier Red Team, research post, October 27, 2026
Anthropic said it reported the work to Google and that 62 CVEs, all Critical or High, had been assigned and patched as of Chrome 156. That 62 is a version-wide tally across Chrome 156 builds, including the large October 21 drop. It is not a second count of the 52 items in the October 27 desktop note, and the two figures should not be added or averaged.
The practical point for a fleet is narrower. A lab model that can string renderer bugs into a Calculator pop is the reason a staged rollout of four fresh critical fixes is a bad week to sit on 156.0.8078.47. The window between a public note and a working exploit is the thing that shrank, even if Google still has not flagged in-the-wild use for this batch.
Gemini 4 Carbon Opens First to Cyber Defenders
Koray Kavukcuoglu, SVP of Google DeepMind and Google’s chief AI architect, introduced Gemini 4 Carbon on October 28 as the company’s strongest model, with a lead on cybersecurity tests. Google said Carbon scored 88.2% on SEC-bench Pro for finding bugs, 54.8% on ExploitGym for exploitation, and 71.0% on Collinear’s CWE-bench for patching, all measured without guardrails. It also ranked first on Artificial Analysis Intelligence Index v5.0 with a score of 59.
CARBON’S PUBLISHED SECURITY SCORES
| Test | Carbon | What Google says it measures |
|---|---|---|
| SEC-bench Pro | 88.2% | Finding real vulnerabilities |
| ExploitGym | 54.8% | Checking that a bug can be exploited |
| CWE-bench | 71.0% | Turning findings into working patches |
| AA Intelligence Index v5.0 | 59 (rank 1) | Broad model quality |
Google’s announcement materials placed GPT-6 Astra second on SEC-bench Pro at 85.4% and Gemini 4 Argon at 83.4%. Argon had launched two weeks earlier. On internal tests, Carbon more than doubled Argon’s success rate at generating secure code. The Chrome team, Kavukcuoglu wrote, reported that Carbon was more than 10X better at patches than another lab’s large frontier model. Google did not name that lab.
Carbon delivers state-of-the-art performance across the cybersecurity defense lifecycle:
🔵 Finding real vulnerabilities: 88.2% on SEC-bench Pro
🔵 Validating risk through exploitation: 54.8% on ExploitGym
🔵 Patching at the root cause: 71.0% on Collinear’s CWE-benchOn… pic.twitter.com/lN61vR1WIB
— Google (@Google) October 28, 2026
Access is the other half of the announcement. Google said it is in the U.S. government’s voluntary pre-release testing process, and that Carbon would go first to a subset of cyber defenders, then to more trusted partners through the Fairwind Program for trusted partners. Developers, businesses, and consumers were told to wait. Introductory API pricing, when it arrives, is $2 per million input tokens and $10 per million output tokens, with cached input at 95% off; after the intro period those rates become $4 and $20.
Artificial Analysis’s intelligence index puts Carbon at $7.76 per task, Claude Opus 5.5’s maximum configuration at $10.04, and GPT-6 Astra’s maximum at $6.25. Those are cost figures, not a claim that Carbon is already hunting Chrome bugs in the wild. The model that is already named in Chrome’s critical-bug credits is Claude, and it is not Google’s.
What Managed Chrome Still Has to Catch
Enterprises need the installer on each device, not the blog post. Google is spreading 156.0.8078.73 and 156.0.8078.74 over days and weeks, and a policy hold, a lagged admin console, or a user who never restarts can leave the four critical bugs in place. On October 29, some Windows and Mac installs still showed 156.0.8078.47 until the About Chrome page was opened, and some Android units were still waiting on Play.
WHAT TO VERIFY ON EACH DEVICE
- Desktop build: Help, then About Google Chrome, should read 156.0.8078.73 or later; 156.0.8078.74 on Windows and Mac counts.
- Android build: 156.0.8078.73 from Google Play, then a relaunch, because store rollout lags the desktop note.
- Admin holds: Chrome browser management policies that pin versions or delay updates will skip this drop until someone lifts them.
- Other Chromium browsers: Edge, Brave, Opera, and the rest ship their own builds on their own clocks, even when the bugs live in shared code.
The About page is the check that actually matters, because the staged wave does not announce itself. Opening that page is what forced several still-unpatched desktops to start the download on October 29. Waiting for a silent background update is how a FileSystem critical reported on October 7 stays live on a work laptop two days after Google published the fix.
Carbon will not help those machines. It is not on the public API yet, and even when Fairwind widens, it is a model for people who already hunt and patch bugs. The fleet problem is still a version string, a restart, and whatever Chromium fork the company actually deploys.
2,910 Chrome CVEs in 2026 Reset Patch Cadence
The 52-fix note looks large only if 2025 is the baseline. A public tracker, stack.watch, counts 2,910 Chrome vulnerabilities in 2026 through late October, against 247 in all of 2025 and 290 in 2024. That is 2,663 more than last year, with an average score of 6.99 against 7.10 in 2025.
CHROME CVE COUNTS BY YEAR
| Year | Chrome CVEs | Average score |
|---|---|---|
| 2026 (through late October) | 2,910 | 6.99 |
| 2025 | 247 | 7.10 |
| 2024 | 290 | 7.71 |
Those yearly totals mix Google’s own finds, outside researchers, and the AI-tagged reports that now sit in the same changelog. Thirty-two of the 52 bugs in the October 27 build are memory corruption, use-after-free, or out-of-bounds writes, which is the same class of flaw Chrome has been shipping patches for all year. A four-week browser train that once absorbed a few dozen CVEs is now absorbing hundreds, then dozens more in a follow-up build, then a research post that says a model already turned four of them into a Calculator demo.
On October 29, some desktops still sat on 156.0.8078.47 until someone opened About Chrome. Carbon, the model Google says writes patches more than 10 times better than a rival, was still limited to a subset of cyber defenders. The four critical bugs Ge filed in October were already in the stable notes, waiting on a restart.
NEWS2 months agoClancy Juror Says the Holdout Never Cited Evidence
NEWS3 months agoThe iPhone 18 Pro Max Splits Apple’s Fall in Two
NEWS3 months agoOpenAI Calls GPT-6 Astra AGI Then Locks Its Exploits
BUSINESS3 months agoTrump Claims Hormuz Control as Oil Holds Near $96
LIFESTYLE3 months agoRussia Closes the Goethe-Institut Berlin Tried to Save
AUTO3 months agoVolkswagen’s Job Cuts Leave Four Plants Without Cars
NEWS3 months agoDyson Puts Nurovi Robots on Sale Without the UV Flagship
BUSINESS2 months agoUS Destroys Five Iranian Tankers as the Shock Hits Jordan













