Breaking

Anthropic Says Unreleased Claude Model Built Five Exploit Chains

October 28, 2026 at 08:32 EDT

Security
Foundation Models

Anthropic says an unreleased internal Claude research model built five working exploit chains across Chrome, Safari, Android, Windows and Linux. The demonstrations ranged from browser compromise and data access on a locked phone to remote root access on a configured Linux server. Anthropic’s report describes tests conducted over several months; it does not name the model used.

OCT 27, 2026 · SECURITY RESEARCH

Exploit chains span browsers and operating systems

An unreleased Claude research model produced five attack chains across widely used software. The demonstrations ranged from browser compromise to remote Linux root access.

Inside the reported Chrome findings

Anthropic counted 62 critical- or high-severity CVEs, including these two reported categories.

Sandbox escapes

Renderer compromise flaws

Five targets, distinct impacts

ChromeCode execution beyond the browser sandbox

SafariCross-site access to browser data

AndroidData extraction from a locked phone

WindowsRemote machine crashes

LinuxRemote command execution as root

Defensive value

Findings were reported to vendors, and Anthropic recommends scanning code during development.

Limits of the disclosure

The model is unnamed, human input is not broken out, and the demonstrations were not reported as real-world attacks.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.