Back to the stories

Anthropic says Claude discovered and exploited critical vulnerabilities in widely used software

Score 10.0,

AI cybersecurity

Building on our coverage from three days ago, Anthropic has now laid out a broader set of findings and fixes. An unreleased Claude research model found attack paths in browsers, a phone, and core operating systems.

Anthropic says it reported the findings to the companies responsible. Google, Apple and Microsoft addressed reported issues, and Linux fixes were added to maintained versions.

Claude linked a Chrome flaw to a way of running code beyond the browser’s protections. In Safari, it found a path to private browser data.

This matters because these products are widely used. One hidden weakness can put many people’s information or systems at risk.

Anthropic says its system sends many AI workers to search in parallel, then checks and ranks their findings. Think of it as a team of investigators following different leads, with each finding checked before it is treated as a real flaw.

The striking part is the range: one research effort found weaknesses in browsers, a phone, and operating systems. But the model is unreleased, not an open tool for anyone to run. Anthropic is expanding access to advanced cyber capabilities for qualifying security professionals, so they can test defenses.

The next question is how to give defenders enough access to find flaws early, without putting the same capability in the wrong hands.