AI 当黑客初显身手：Claude 自主挖洞写利用，53 分钟拿下 Linux 最高权限 Anthropic 10 月 27 日甩出一份让安全圈侧目的报告：一个未公开的 Claude 内部研究模型，只拿到初始提示，便自行定位漏洞、编写利用代

AI first showed its skills as a hacker: Claude independently dug holes and wrote exploits, and obtained the highest permissions on Linux in 53 minutes. On October 27, Anthropic released a report that raised eyebrows in the security circle: an undisclosed internal research model of Claude. After only receiving initial prompts, it located vulnerabilities and wrote exploit codes on its own, uncovering flaws on Chrome, Safari, Android, Windows and Linux, and stringing them together into 5 An attack chain. All vulnerabilities have been reported to the manufacturer and fixed. The most shocking thing is that the Linux case model discovered a zero-click remote code execution vulnerability (CVE-2026-98432) in the kernel network stack. The target is a server running a multi-path TCP listener - the attacker only sends data packets and can get the root shell without any credentials. Anthropic actually measured that the entire chain took 53 minutes, of which 48 minutes were used to leak the kernel pointer and the remaining 5 minutes were used to execute the command. Android is equally outrageous: with the help of USB flaws, attackers can extract local data such as photos even if the Pixel 9 has a locked screen. Chrome contributed 62 high-risk CVEs in one place. Chrome is the largest. The report submitted by Claude to Google revealed 62 critical or high-risk CVEs at once, including 13 sandbox escape and 21 controllable rendering process vulnerabilities. The patch was released with Chrome 156; the calculator popped up within seconds of loading the macOS proof-of-concept page. The Safari chain bypasses the same-origin policy, allowing malicious web pages to read history, cookies, logged-in emails, and banking data with one click. Apple fixed it on the same day. On the Windows side, the model discovered an HTTP/3 (QUIC) kernel out-of-bounds read vulnerability (CVE-2026-33096, score 7.5), which can remotely crash the server. Microsoft has issued a patch in April. via AI News (author: AI Base)