Cathedral’s researchers uncover weaknesses in third-party business software, highlighting the cybersecurity risks facing companies developing advanced artificial intelligence
San Francisco, 26 October 2026 – A cybersecurity startup called Cathedral has reportedly gained access to parts of Anthropic’s internal business systems in less than 12 hours, exposing weaknesses in software used to manage everyday corporate operations. The incident highlights a growing challenge for artificial intelligence companies as they work to protect sensitive information while relying on multiple external technology providers.
According to a report by The Washington Post, Cathedral’s researchers discovered ways to reach certain internal systems while participating in Anthropic’s security testing program. The researchers reportedly used AI agents to identify weaknesses in third-party business applications, which eventually provided access to limited internal information.
The information they could reach reportedly included meeting transcripts, employee salary and job offer details, and internal Slack discussions about advanced AI models. Some customer and partner contact records were also involved.
Anthropic confirmed that the researchers accessed a limited set of its sales, human resources, and information technology systems. However, the company said the researchers did not access customer conversations with Claude, responses generated by the AI assistant, source code, or model weights. Model weights are the numerical values that help determine how a trained AI model behaves.
The incident draws attention to a security problem that extends beyond AI models themselves. Large technology companies often depend on third-party applications for recruitment, communication, sales, administration, and other routine activities. If these applications contain vulnerabilities, attackers may find ways to reach information stored within connected corporate systems.
For companies working with advanced AI, the consequences could be significant. Internal discussions, employee records, and research related to new technologies can contain confidential information that businesses need to protect.
Cathedral reported its findings to Anthropic through the company’s vulnerability disclosure program, which allows external researchers to identify security weaknesses and report them. Anthropic said it disabled the affected access routes and worked with software vendors to address the vulnerabilities. The company also informed federal officials.
Anthropic said its monitoring systems had detected some activity before the findings were disclosed and that it had no evidence of other parties exploiting the vulnerabilities.
The company also disputed aspects of Cathedral’s account, arguing that the startup overstated the national security implications and did not fully comply with the rules governing its security testing.
The incident demonstrates how AI-powered cybersecurity tools can help researchers identify weaknesses faster than traditional methods. At the same time, similar capabilities could create risks if used without authorization or appropriate safeguards.
For AI developers, protecting customer information and intellectual property requires more than securing their own software. They must also examine how external applications connect to internal networks, restrict access to sensitive records, and monitor unusual activity.
Security experts have increasingly highlighted third-party software as a potential entry point for cyberattacks. As companies adopt more AI agents and automated tools, controlling what these systems can access will become equally important.
The Cathedral incident serves as a reminder that even organizations developing sophisticated AI technology can face vulnerabilities in ordinary business systems. Stronger vendor security checks, tighter access controls, continuous monitoring, and responsible vulnerability disclosure will remain essential to protecting sensitive corporate information as AI adoption expands.