Anthropic’s research report details how an unreleased internal Claude model autonomously discovered and chained multiple vulnerabilities across Chrome, Safari, Android, Windows, and Linux into working exploits. The model combined four flaws in Chrome’s V8 engine and sandbox protections for code execution outside the sandbox. It bypassed Safari’s same-origin policy to extract history and cookies. On Android, it chained USB and pairing authorization flaws to unlock phone data. The Windows exploit targeted a QUIC memory safety flaw, and on Linux it achieved kernel RCE through a multipath TCP race condition.
Autonomously weaponizing bug chains requires discovering multiple vulnerabilities and reasoning about how they compose into end-to-end exploits. That’s a different threat level than finding individual flaws.
The guardrails are important: this happened in lab conditions with an unreleased model, not deployed. Anthropic disclosed to vendors first, and all patches landed before publication. The disclosure process worked correctly.
Sources: Discovering and exploiting vulnerabilities in critical, real-world software with Claude
Coverage: Claude Model Builds Exploits Across Major Browsers & OSes
Related on this blog: [AUTO] The Exploit Window Collapsed • [AUTO] When Open Weights Enable Exploit Generation • [AUTO] Claude Mythos autonomously chains zero-days into exploits