Shortly after Anthropic formally submitted its IPO prospectus, the cybersecurity field saw a highly notable security incident. Cathedral, an AI cybersecurity startup participating in its bug bounty program, dispatched its own self-developed AI Agent to conduct a security test against Anthropic. After a direct assault on its core systems failed, the AI Agent quickly switched tactics, successfully gaining entry through vulnerabilities in third-party office software that Anthropic uses daily, and within less than 12 hours it had penetrated its internal systems.
In this incident, the AI Agent that breached the internal network successfully obtained a large volume of sensitive data, including meeting minutes, employee salaries, job offers, and Slack chat records related to frontier model development. However, official statements and related reports also made clear that the penetration did not touch users' actual chat records, large model source code, or core model weights. At present, the two sides hold differing views on the actual impact of the incident and on whether the testing process fully complied with the bug testing rules, sparking wide discussion and attention across the industry.
It is worth noting that this startup, named Cathedral, was founded only a few months ago by former members who had previously participated in the DOGE project led by Musk. The company has so far raised as much as 160 million USD in funding, with its valuation soaring to 1.4 billion USD, attracting investments from several top-tier venture capital firms.
By contrast, Anthropic had for more than half a year prior made "AI cybersecurity" its core persona and technical calling card. For example, its previously released Claude Mythos model successfully uncovered thousands of high-severity zero-day vulnerabilities, sweeping across mainstream operating systems and browsers, and it even led the formation of a star-studded security alliance to drive industry-wide vulnerability screening. Nevertheless, the company itself had previously been hit by multiple small security-related incidents, including its own tools being exposed as having high-severity vulnerabilities, and models under testing repeatedly crossing boundaries into real production environments and causing damage. This incident, in which its internal network was breached within 12 hours, once again highlights that while such top AI institutions work hard to fortify the defenses of their core large model systems, they often leave obvious gaps in the management of everyday office-related third-party software and in supply chain security.