Live monitoring Updated 72-hour alert levelPriority Editor: Frédéric LENQUETTE
CYBERIACKCybersecurity & AI intelligence · by Frédéric LENQUETTE
AI & securityPriority

Anthropic says one of its AIs designed attacks on Chrome, Safari, Windows and Linux

On October 27, Anthropic's security team published five attack chains produced by an internal model. The flaws are patched: here are the versions to install, and what this announcement does not yet prove.

Translated from French by AI. The original French article is the reference version.

Article illustration : Anthropic says one of its AIs designed attacks on Chrome, Safari, Windows and Linux
AI-generated illustration

One click on a link in Chrome, and a Mac's calculator opens by itself. It is proof that foreign code was able to run on the machine. This demonstration appears in a post published on October 27 by Anthropic's Frontier Red Team, the group tasked with testing the dangerous capabilities of its models. According to the company, this attack, like four others, was designed by an internal research model, unreleased and unnamed, "with little human intervention."

The five chains target Chrome, Safari, Android Pixel 9 phones, Windows and the Linux kernel. Anthropic says it reported all the flaws to the vendors. According to the company, all of them were patched before publication.

Five attacks, from the most harmless to the most serious

  • Chrome: four chained vulnerabilities allow code execution after a single click. According to Anthropic, its reports to Google led to 62 vulnerability identifiers (CVEs), all rated critical or high and fixed in Chrome 156 at the latest.
  • Safari: a one-click chain bypasses the rule that isolates websites from one another. Apple has released a patch.
  • Pixel 9: by plugging the phone in over USB, photos could be extracted without unlocking it. Google says it has "addressed" the problem.
  • Windows: the flaw CVE-2026-33096 could remotely crash servers that respond over HTTP/3. Microsoft has fixed it.
  • Linux: the most serious. The flaw CVE-2026-98432 is in the kernel's network stack. It allows full takeover of a server that accepts Multipath TCP connections, with no action by its user. According to Anthropic, the demonstration takes about 53 minutes to succeed.

What administrators should check

Good news: none of this is an open vulnerability today, provided you are up to date. Update Chrome to version 156 or later, and apply the latest Apple and Pixel updates. On Windows, the fix dates from the April 2026 updates: a server that missed them remains exposed. On Linux, Anthropic says the fix has been backported to all maintained stable branches. The site Cyber Kendra cites kernels 6.1.190, 6.6.159 and 6.12.113, among others. The same site notes that Red Hat deferred the fix for RHEL 10 and that Ubuntu 26.04 was still under evaluation: check with your distribution.

What the announcement does not yet prove

This account comes from Anthropic, which has an interest in demonstrating the power of its models ahead of its stock market listing. Several points cannot be verified from the outside. The model is not accessible. Anthropic acknowledges that each chain started from a human prompt and that, for Windows, researchers supplied the program to analyze. Chrome's release notes credit Anthropic for some of the flaws. But the real share of autonomy the AI had in each attack is not established. And specialized outlets point out that the total of 62 Chrome flaws is not fully corroborated by Google's release notes.

Even so, the message is close to that of XBOW, which hacked Chrome with an OpenAI model, or of the Linux flaw attributed to a Tencent AI. Finding and chaining vulnerabilities is no longer reserved for a few elite teams. Anthropic draws one instruction from this: analyzing your own code "is more important than ever." It is also widening access for vetted professionals to less restricted versions of its models. We will have to watch who really benefits, defenders or attackers.

Update of November 13, 2026

The Linux flaw CVE-2026-98432, exploited in one of the five attack chains Anthropic presented on October 27, is now patched in the two distributions that were missing. Red Hat released patched kernels on November 1 for RHEL 8 (advisory RHSA-2026:93414, version 4.18.0-553.182.1.el8_10), RHEL 9 (advisory RHSA-2026:93412, version 5.14.0-687.64.1.el9_8) and RHEL 10 (advisory RHSA-2026:93413, version 6.12.0-211.72.1.el10_2). Ubuntu states on its tracker page, updated on November 11, that the kernel for version 26.04 LTS is patched (7.0.0-43.43), as are those for 22.04, 24.04 and 26.10.

Fixes are still being prepared for older Ubuntu versions (16.04, 18.04 and 20.04). In all cases, the machine must be rebooted for the patched kernel to take effect.

Method: AI-assisted monitoring, verified sources, article reviewed and approved by Frédéric LENQUETTE. Our editorial policy.

CVE entries: CVE-2026-33096 CVE-2026-98432

Vendor page: Linux · 32 exploited vulnerabilities

Share this article