A cybersecurity startup has exposed a potentially serious weakness in the technology infrastructure of Anthropic, the company behind Claude, after its AI agents reportedly found routes into internal business systems in less than 12 hours. The findings have raised questions about how effectively leading AI companies are securing the software and integrations that sit around their core technology.
The startup, Cathedral, conducted the exercise under Anthropic’s programme for external security researchers to identify vulnerabilities. According to a draft report reviewed by The Washington Post, Cathedral’s AI agents found weaknesses in third-party software used for administrative and sales operations, opening paths towards sensitive internal information.
The reported access included meeting transcripts, employee salary and job-offer details, and internal Slack discussions concerning Anthropic’s advanced AI models. However, Anthropic has disputed the broader scope of the reported exposure, confirming that researchers accessed a limited set of its internal sales, human resources and IT systems.
Third-Party Software Emerges as the Weak Link
The incident is notable because the reported entry points were not vulnerabilities in Claude itself. Instead, Cathedral identified weaknesses in business applications connected to Anthropic’s internal operations.
These applications support routine enterprise functions, including administration, sales and employee management. Yet their connections to corporate systems can create access routes to sensitive records if security controls, permissions or software integrations are not properly managed.
Anthropic said the information accessible to researchers included business contact details for some customers and partners, along with basic user-account records. The company also stated that customer conversations with Claude, the model’s responses, proprietary source code and trained model weights were not accessed.
This distinction matters. While the incident did not involve a reported compromise of Anthropic’s core AI models, it highlights how weaknesses in connected enterprise applications can create risks for organisations handling valuable intellectual property and confidential business information.
The affected software vendors have not been publicly identified in the available reporting.
AI Agents Bring Speed to Security Testing
The reported 12-hour timeline is one of the most striking aspects of the incident. Cathedral’s findings suggest that AI agents can help researchers investigate complex software environments and identify potential access routes at considerable speed.
For security teams, the same capabilities could strengthen vulnerability assessments, automate routine checks and help uncover weaknesses before malicious actors exploit them. However, faster discovery also puts pressure on organisations to improve monitoring, access controls and incident response.
The broader concern is not that AI agents can automatically penetrate any corporate network. Rather, their ability to perform multi-step tasks can help connect individual weaknesses into more consequential access paths, particularly where third-party applications have broad permissions or insufficient safeguards.
For enterprises increasingly deploying AI agents across business operations, the incident reinforces the need to understand which systems these agents and connected applications can reach, what information they can access, and how their activity is monitored.
Anthropic Says It Detected the Activity
Anthropic said its monitoring systems flagged some of Cathedral’s activity before the startup reported its findings. The company said it disabled the researchers’ access on the same day and temporarily disabled the affected software while vendors addressed the vulnerabilities.
Anthropic also said it had found no evidence that anyone other than Cathedral had exploited the weaknesses.
The two sides differ over the interpretation and handling of the exercise. Anthropic said Cathedral overstated the national security implications and did not fully comply with the programme’s disclosure requirements. Cathedral’s draft report maintained that data access had been kept to the minimum necessary for the research.
The incident therefore needs to be viewed in its proper context: it emerged from an authorised security testing exercise, rather than a confirmed criminal attack. At the same time, such exercises can reveal weaknesses that organisations need to resolve before they are discovered by actors with malicious intent.
A Wider Warning for the Technology Industry
Cathedral was founded by former members of the US Department of Government Efficiency and has attracted significant investor backing. The startup reportedly raised $160 million at a valuation of $1.4 billion in July 2026, reflecting growing interest in AI-powered cybersecurity capabilities.
For the wider enterprise technology industry, the case underscores a fundamental challenge. Organisations often invest heavily in securing their own infrastructure while relying on external software providers and interconnected applications to manage critical business functions. Each connection needs appropriate access restrictions, oversight and monitoring.
As AI agents become more capable, cybersecurity strategies will need to account for how quickly these systems can identify and follow potential routes through connected software environments. That means strengthening vendor risk assessments, reviewing application permissions, monitoring unusual activity and ensuring vulnerabilities are addressed promptly.
Cathedral’s findings offer a timely warning for businesses building or adopting AI technologies: protecting the core model is only one part of the security equation. The applications, integrations and internal systems surrounding it can be just as important in determining how well an organisation protects its most sensitive information.
𝐒𝐭𝐚𝐲 𝐢𝐧𝐟𝐨𝐫𝐦𝐞𝐝 𝐰𝐢𝐭𝐡 𝐨𝐮𝐫 𝐥𝐚𝐭𝐞𝐬𝐭 𝐮𝐩𝐝𝐚𝐭𝐞𝐬 𝐛𝐲 𝐣𝐨𝐢𝐧𝐢𝐧𝐠 𝐭𝐡𝐞 WhatsApp Channel now! 👈📲
𝑭𝒐𝒍𝒍𝒐𝒘 𝑶𝒖𝒓 𝑺𝒐𝒄𝒊𝒂𝒍 𝑴𝒆𝒅𝒊𝒂 𝑷𝒂𝒈𝒆𝐬 👉 Facebook, LinkedIn, Twitter, Instagram