RC RANDOM CHAOS

Claude Chains Critical Vulnerabilities Into Working Exploits

· via Dark Reading

Original source

Claude Model Builds Exploits Across Major Browsers & OSes

Dark Reading →

An unreleased Claude model discovered and chained critical vulnerabilities in Chrome, Safari, Android, Windows, and Linux into five functional exploits. The model combined four Chrome flaws to escape the browser sandbox and achieve remote code execution on macOS. For Safari, it bypassed the same-origin policy through a multi-step attack on WebKit and JavaScriptCore, potentially exposing authenticated sessions for email and banking. On mobile, the model extracted data from a locked Pixel 9 by chaining Android vulnerabilities. Linux faced a zero-click root-level remote code execution threat via a kernel networking flaw. All vendors have since patched the issues.

The demonstration marks a significant shift in attack sophistication: what once demanded elite reverse engineering teams can now emerge from a single model run. Anthropic is expanding its Cyber Verification Program to provide security professionals access to its most powerful models, betting that defenders need equivalent capabilities to find and patch vulnerabilities before attackers exploit them.

Security researchers acknowledge the findings while noting important constraints. The exploits succeeded only under specific conditions—the Android attack required a stolen phone and USB cable; the Linux exploit needed an exposed network listener. Existing mitigations like ASLR, stack canaries, and sandboxing substantially reduce AI-discovered vulnerability success rates in real deployments. However, experts warn the gap between proof-of-concept and reliable field attacks is narrowing from one requiring specialized expertise to one mainly requiring adaptation time and resources. This escalates risk for less-scrutinized infrastructure: embedded firmware, IP cameras, and office routers with minimal security review.

Read the full article

Continue reading at Dark Reading →

This is an AI-generated summary. Read the original for the full story.