Chrome vulnerabilities: an Anthropic model found 62 bugs, all now fixed

What happened. On 27 October Anthropic published a report from its Frontier Red Team. An internal model the company has not released looked for bugs in Chrome, Safari, Android, Windows and the Linux kernel. It chained the bugs it found into five attacks and showed they work. Every flaw went to the maker first, and all are patched. None were seen in real attacks.

Chrome. Anthropic says 62 flaws received CVE numbers, all rated critical or high. They include 13 sandbox escapes, meaning ways out of the isolated zone where the browser keeps website code, and 21 flaws that can take over the page rendering process. The chain could launch a program on a Mac with one click, in a browser with default settings. All are fixed as of Chrome 156.

What is in Chrome 156. Google moved Chrome 156 to the stable channel on 20 October: version 156.0.8078.47 for Linux and 156.0.8078.47/.48 for Windows and Mac. The update has 353 security fixes. Google credits some of them to "Xinyang Ge (Anthropic), assisted by Claude", meaning an Anthropic researcher working with Claude.

Other systems. In Safari the chain broke the same-origin policy, the rule that stops one site reading another site's data. The Hacker News says this is fixed in iOS and iPadOS 26.5.2. On a Google Pixel 9 the model got past the lock screen and pulled photos off a locked phone, and Google has fixed it. In Windows it found a crash that can be triggered remotely in QUIC and HTTP/3 handling (CVE-2026-33096). In Linux it found a way to get root over the network with no login and no user action, through multipath TCP (CVE-2026-98432). That was fixed upstream and backported to all maintained stable kernel branches.

Disclosure rules. Makers get 90 days to fix a flaw. Anthropic then waits 45 more days before sharing details beyond a public advisory. The company also widened its Cyber Verification Program, which gives vetted security professionals access to models with fewer limits for defensive work.

What is unclear. The Hacker News notes that how independently the model worked cannot be checked from outside, because Anthropic does not say where people helped. It also says the public CVE records for Windows and Linux do not describe the bugs exactly the way the report does. The report does not say whether Anthropic will release the model.

What to do. Update Chrome: menu, Help, About Google Chrome. The version should start with 156 or higher. Update your iPhone and iPad. Anthropic's advice to defenders: scanning your own code matters more than ever.

Fatos confirmados

• On 27 October 2026 Anthropic's Frontier Red Team published a report on finding and exploiting flaws in Chrome, Safari, Android, Windows and Linux with an unreleased internal model.^[1][3]
• In Chrome, 62 flaws received CVE numbers and are patched as of Chrome 156; all are rated critical or high, including 13 sandbox escapes and 21 renderer compromise flaws.^[1][3]
• The Chrome chain could launch a program on macOS with one click in a default-configuration browser.^[1]
• Chrome 156 reached the stable channel on 20 October 2026 (156.0.8078.47 for Linux, 156.0.8078.47/.48 for Windows and Mac) with 353 security fixes.^[2]
• Chrome 156 release notes credit several fixes as "Reported by Xinyang Ge (Anthropic), assisted by Claude".^[2]
• The Safari chain bypassed the same-origin policy; The Hacker News says it is fixed in iOS and iPadOS 26.5.2.^[1][3]
• For Linux, a zero-click remote root attack through multipath TCP (CVE-2026-98432) was fixed upstream and in stable kernel branches.^[1]
• For Windows, a remotely triggerable crash in QUIC and HTTP/3 handling was assigned CVE-2026-33096.^[1][3]
• All flaws were reported to the makers and patched; none were seen in real attacks.^[3]
• Anthropic's disclosure policy is 90 days to patch plus 45 more days before details beyond a public advisory.^[1][3]
• Anthropic widened its Cyber Verification Program for vetted security professionals.^[1]
• The Hacker News says the model's autonomy cannot be independently verified and the Windows and Linux CVE records do not fully match Anthropic's descriptions.^[3]

Implicações

• ConfirmadoAll 62 Chrome flaws in the report are fixed in Chrome 156, so updating matters.
• Nossa leituraAI makes finding complex attack chains faster and cheaper, so browser and system updates become even more important.
• Nossa leituraCompanies that do not scan their own code with tools like these risk attackers finding the bugs first.

Ainda não se sabe

• How much the model did on its own and where people helped.
• Whether and when Anthropic will release this model.
• Whether others already have models with similar abilities.

Quem é afetado

Global. Anyone who uses Chrome, an iPhone, Android, Windows or Linux, and anyone following AI and cybersecurity

Fontes

1. Anthropic: Discovering and exploiting vulnerabilities in critical, real-world software with Claude 27 de outubro de 2026 · oficial
2. Google Chrome Releases: Stable Channel Update for Desktop 20 de outubro de 2026 · oficial
3. The Hacker News: Anthropic's Unreleased AI Model Chains Flaws Into Working Chrome, Safari, Windows and Linux Exploits 28 de outubro de 2026 · mídia