What happened
According to a report by IT之家 citing The Washington Post, the cybersecurity company Cathedral said its researchers managed to enter Anthropic's internal systems by exploiting vulnerabilities in third-party commercial software, achieving the objective in less than twelve hours through the use of artificial intelligence agents. The news was released on October 26 and picked up the following day.
Anthropic confirmed the incident, specifying that the access involved a limited perimeter: sales, human resources, and IT systems. The company ruled out consequences for customer-Claude conversations, model responses, source code, and model weights—that is, the parameters that define how the AI operates.
The context of the test and the information exposed
Cathedral, founded by former members of the U.S. government efficiency department, conducted the test as part of Anthropic's external security research program. According to a draft report not yet published, cited by the source, the researchers gained access to meeting minutes, information on employee salaries and hiring, and Slack messages discussing advanced AI models.
Anthropic said it had identified some of the relevant activity through its own monitoring systems even before Cathedral presented the report, revoking the researchers' access the same day. The company also suspended use of the third-party software involved until the vendor confirmed the vulnerabilities had been fixed, and reported the incident to the relevant U.S. federal authorities.
The matter is significant because it involves a leading AI model provider and because the described method combines automated agents and flaws in common enterprise software. However, details remain incomplete: the name of the software vendor and the specific characteristics of the vulnerabilities have not been disclosed.
What remains to be verified
The account is based on statements by Cathedral, reported in a not-yet-published report, and on Anthropic's confirmation. No independent elements are available to quantify the scope of the data consulted, the actual duration of access, or the effectiveness of the measures adopted.
Anthropic maintains that there is no evidence that parties other than Cathedral used the vulnerabilities; this is the company's assessment, not a fact verified by third parties. It also remains unclear how the vendor will fix the defects and whether the case will have regulatory consequences, given the report to federal authorities.
