AI Agents Identify Security Weaknesses at Anthropic

Cathedral, an AI-focused cybersecurity startup, reportedly gained access to certain internal systems at Anthropic in less than 12 hours, raising fresh questions about how businesses secure sensitive information in an era of increasingly capable AI agents.

Cathedral's researchers used AI agents to identify vulnerabilities in third-party business software connected to Anthropic's sales, human resources and IT operations. The researchers reportedly accessed employee information, salary details, job offers, meeting transcripts and internal messages discussing the company's advanced AI models.

The incident was discovered during security testing conducted through Anthropic's vulnerability disclosure program, which allows outside researchers to investigate security weaknesses and report their findings.

Third-Party Software Was the Entry Point

The reported access did not result from directly breaking into Anthropic's AI models. Instead, researchers exploited weaknesses in external business applications used for administrative and operational tasks.

Such applications can hold sensitive employee and business information or connect to internal company systems. If they contain vulnerabilities or are configured insecurely, they may provide an entry point into an organization's wider technology environment.

The incident highlights the importance of securing the entire software ecosystem, rather than focusing exclusively on a company's own applications and infrastructure.

What Information Was Exposed?

Reports indicate that Cathedral's researchers reached internal business records and communications, including employee compensation information, employment offers and Slack discussions about Anthropic's AI development.

Anthropic confirmed that a limited set of sales, HR and IT systems had been accessed. However, the company said its customer conversations with Claude, proprietary source code and AI model weights were not accessed.

Anthropic also said it disabled the affected software and found no evidence that other parties had exploited the vulnerabilities. The company disputed aspects of Cathedral's account, including its handling of the security testing process, while Cathedral said it kept data access to a minimum.

These distinctions are important: access to internal business systems can create significant privacy and security risks, but it does not automatically mean that a company's AI models or customer conversations have been compromised.

Why AI-Powered Cybersecurity Matters

AI agents can perform multistep tasks with less continuous human direction than traditional software tools. In cybersecurity, this capability can help researchers search for weaknesses, investigate systems and identify potential attack paths more quickly.

The same capabilities also create challenges for organizations. Security teams must consider how automated tools interact with connected applications, what permissions they receive and whether their actions can be monitored or contained.

The Cathedral incident illustrates how vulnerabilities in ordinary enterprise software can become important security concerns, even at companies with substantial technical expertise.

Lessons for Enterprise Security Teams

The reported incident offers several lessons for organizations that rely on interconnected business applications:

  • Audit third-party software: Review applications that store sensitive information or connect to internal systems.

  • Apply least-privilege access: Limit employee accounts, applications and automated agents to the information and functions they genuinely require.

  • Monitor unusual activity: Use logging and security monitoring to identify unexpected access to internal records and communications.

  • Strengthen vulnerability management: Establish clear procedures for reporting, investigating and fixing security weaknesses.

  • Govern AI agents carefully: Set boundaries on the systems they can access, the actions they can take and the data they can retrieve.

The Road Ahead for AI and Cybersecurity

As AI agents become more capable, organizations will need to adapt their security strategies to account for both the opportunities and risks of automation. AI-powered testing can help uncover weaknesses, but its effectiveness depends on clear authorization, controlled access and responsible disclosure.

For enterprises, the broader message is that cybersecurity cannot stop at the perimeter of their own technology. Third-party applications, connected services and automated agents all form part of the security environment.

The Cathedral incident is a reminder that even highly sophisticated AI companies must continuously evaluate the systems surrounding their core technology. Protecting business data requires attention to the entire digital ecosystem, not just the AI models at its center.