Anthropic's Frontier Red Team: unreleased Claude built exploits for five OSes

The disclosure comes from Anthropic's Frontier Red Team, the group inside the company that stress-tests its own models for dangerous capability. Rather than a hypothetical benchmark, the team reports end-to-end chains: a zero-click remote-root path on Linux tracked as CVE-2026-98432 with a CVSS score of 9.8, and a one-click Chrome escape that combines four V8 and sandbox bugs patched in Chrome 156.

The other chains stretch across a Safari same-origin break through WebKit's Gigacage, a locked-Pixel-9 bypass over USB, and a Windows QUIC crash (CVE-2026-33096). The body notes that all flaws were reported to vendors ahead of patches, which places the work in the coordinated-disclosure track rather than a live attack report.

The paper's framing is the part likely to travel: frontier models can now "prove" novel vulnerabilities by producing working exploits with little human involvement. That shifts the question for defenders from whether a model can describe a weakness to whether it can demonstrate one — a distinction the body draws directly, and one that may put pressure on how quickly vendors ship fixes.