The Hacker Newsinfo@thehackernews.com (The Hacker News)
Anthropic AI model chains flaws into browser and OS exploits
Part of the CVE-2026-33096 story · since 28 Oct
- Wed 28 Oct, 09:02 UTC. The Hacker News, first report. Anthropic AI model chains flaws into browser and OS exploits. This article, 1st of 1.
Anthropic’s Frontier Red Team has disclosed an unreleased internal AI model that reportedly chained together multiple separate browser and OS flaws to produce working exploit chains across Chrome, Safari, Windows and Linux (the Android chain was noted but lacks a public CVE). The research, published by The Hacker News citing Anthropic, describes five chains built after a prompt in a lab, with each chain exploiting distinct vulnerabilities to move from initial access to functionally privileged outcomes.
Anthropic says all flaws were reported to vendors under coordinated disclosure, and patches are in place or documented. One chain allegedly demonstrated running code outside the Chrome sandbox from a single page load on stock macOS Chrome, while others enabled memory reads/writes, or remote kernel effects, depending on platform. The company emphasises that the PoC exploits were lab-created and not observed in live attacks.
Platform-specific details and evidence include: Chrome chain leveraging four flaws in sequence, including a V8 memory-safety issue, a sandbox escape and a CacheStorage race, culminating in a FedCM use-after-free that runs outside the sandbox; patches are linked to Chrome 156 (29 October) with Google crediting Xinyang Ge of Anthropic, assisted by Claude, for fixes on some advisories.
Safari chain could read/write memory and disable same-origin policy without code execution, fixed in iOS/iPadOS 26.5.2 (credit to Milad Nasr and Nicholas Carlini with Claude). Windows chain involves an out-of-bounds read tied to CVE-2026-33096, patched April 2026; Linux chain enables root execution over the network with a multipath TCP listener, linked to CVE-2026-98432, patched October 15.
Anthropic states it documented each flaw and points defenders to its Cyber Verification Program for access to Claude Opus 5.5, Sonnet 5.5 and Mythos 5.1. The U.S. government’s vulnerability catalog had not marked the Windows flaw as exploited at the time of reporting.
How the story developed
1 article from 1 outlet since 28 Oct. Still developing; the latest report was 2 hours ago. This was the first report.
- Wed 28 Oct, 09:02 UTC. The Hacker News, first report. Anthropic AI model chains flaws into browser and OS exploits. This article, 1st of 1.
- CVE not in KEV
- No CVE
- Larger: an outlet’s first report
- Coverage over time
Show all 1 articles as a listHide the list
- Wed 28 OctDay 01
- Anthropic AI model chains flaws into browser and OS exploits09:02 · The Hacker News · First report · This article, 1st of 1