On October 27, Anthropic dropped a report that turned heads in the security world: an undisclosed internal Claude research model, given only an initial prompt, independently located vulnerabilities, wrote exploit code, found flaws in Chrome, Safari, Android, Windows, and Linux, and chained them into 5 attack chains. All vulnerabilities were reported to the vendors and fixed.
The most alarming case was Linux
The model found a zero-click remote code execution vulnerability (CVE-2026-98432) in the kernel network stack, targeting a server running a multipath TCP listener—an attacker could obtain a root shell just by sending packets, without any credentials. Anthropic's testing showed the entire chain took 53 minutes: 48 minutes to leak a kernel pointer, and the remaining 5 minutes to execute commands. Android was just as wild: using a USB flaw, an attacker with physical access could pull local data such as photos even from a locked Pixel 9.
Chrome alone contributed 62 high-risk CVEs
The largest was Chrome. The report Claude submitted to Google yielded 62 critical or high-severity CVEs at once, including 13 sandbox escapes and 21 renderer-process control vulnerabilities; patches landed with Chrome 156. Its macOS proof-of-concept page popped up Calculator within seconds of loading. The Safari chain bypassed the same-origin policy, allowing a malicious webpage to read history, cookies, and email and banking data from logged-in sessions with a single click; Apple fixed it the same day. On Windows, the model found an out-of-bounds read vulnerability in the HTTP/3 (QUIC) kernel (CVE-2026-33096, score 7.5) that could remotely crash servers; Microsoft released a patch in April.