Anthropic’s AI research uncovers flaws across major platforms

Anthropic has disclosed findings from its AI vulnerability research, detailing five attack chains that led to remote code execution and other impacts across popular platforms. Over several months, an unreleased internal model identified flaws in Chrome, Safari, Android, Windows, and Linux, with exploits that could affect billions of users. The researchers say the targets were chosen for their broad reach and potential impact, and that all five chains are now fixed under a coordinated disclosure process.

The Chrome chain produced a one-click exploit that escaped the browser sandbox, with Google patching 62 CVEs from that work—all rated critical or high. A Safari chain bypassed the same-origin policy in a data-only fashion that did not require code execution, and Apple fixed the issue promptly. In Android, Claude demonstrated a method to pull photos from a locked device, which Google has addressed.

For Windows, a flaw in the kernel’s QUIC code allowed remote server crashes; Microsoft fixed it as CVE-2026-33096, with Anthropic estimating millions of such servers publicly accessible. The Linux chain involved a zero-click remote code execution bug in the kernel affecting servers with a multipath TCP listener; CVE-2026-98432 has been patched and backported to all stable branches.

Anthropic states all flaws were disclosed under coordinated disclosure, granting vendors 90 days to patch and a further 45 days before more details emerged.