Claude built its own exploit for a Linux flaw in 53 minutes

What happened

An unreleased Anthropic model autonomously developed a working exploit for CVE-2026-98432, a Linux vulnerability that grants remote root access. According to the company's demo, the process was completed in 53 minutes.

Why it matters

An AI model turning a real vulnerability into a weaponized exploit without human intervention puts direct pressure on patching timelines. If exploits can be generated automatically the moment a flaw is disclosed, organizations' patching windows may shrink to hours.

Details

• Vulnerability: CVE-2026-98432, granting remote root access.
• Model: an unreleased Anthropic version.
• Time: 53 minutes.
• Method: autonomous exploit development; a demo.

Context

AI models' ability to find and exploit security flaws is being tested frequently. This demo shows models can go beyond writing code to assembling end-to-end security chains, making patch prioritization and automated detection more critical for defenders.