马斯克前员工创立的AI初创公司仅用 12 小时攻破Anthropic办公系统 在Anthropic正式递交招股书后不久，网络安全领域迎来了一起备受瞩目的安全事件

An AI startup founded by Musk's former employee took only 12 hours to break into the Anthropic office system. Shortly after Anthropic officially submitted its prospectus, the network security field ushered in a high-profile security incident. Cathedral, an AI network security startup that participated in its bug bounty program, sent its self-developed AI Agent to launch a security test on Anthropic. After failing to attack its core system head-on, the AI ​​Agent quickly changed its strategy and successfully penetrated through the vulnerability of the third-party office software used by Anthropic daily, and penetrated into its internal system in less than 12 hours. In this incident, the AI ​​Agent that broke into the intranet successfully obtained a large amount of sensitive data including meeting minutes, employee salaries, employment offers, and Slack chat records related to cutting-edge model development. However, official and relevant notifications also clearly pointed out that this penetration did not touch the user’s actual chat history, large model source code, and core model weights. At present, both parties have different opinions on the actual impact of the incident and whether the testing process fully complies with the vulnerability testing rules, which has triggered extensive discussion and concern in the industry. It is worth mentioning that this startup called Cathedral was established only a few months ago and was founded by former members of the DOGE project led by Musk. The company has received up to US$160 million in financing, and its valuation has soared to US$1.4 billion, attracting capital injections from a number of top venture capital institutions. On the other hand, Anthropic has been using "AI network security" as its core personality and technology business card for more than half a year. For example, its previously launched Claude Mythos model has successfully discovered thousands of high-risk zero-day vulnerabilities, swept across mainstream operating systems and browsers, and even led the formation of a luxurious security alliance to promote industry-wide vulnerability investigation. However, the company itself has repeatedly exposed security-related episodes before, including the revelation that its tools have high-risk vulnerabilities, and the models under testing have repeatedly crossed the line and broken into the real production environment to cause damage. This 12-hour intranet breach once again highlights that while such top AI institutions are striving to build a strong defense line for core large model systems, they are often prone to leaving obvious omissions in third-party software management and supply chain security related to daily office work. via AI News (author: AI Base)