AI Agents Reached Anthropic’s Business Systems in Under 12 Hours

AI agents found ways into Anthropic's internal systems in under 12 hours, security startup says

Security startup Cathedral says autonomous AI agents found access paths into Anthropic's internal business systems in under 12 hours via sales, HR and IT software rather than the Claude model itself.

The test was conducted under Anthropic's authorized vulnerability disclosure program, an arrangement that lets outside researchers test a company's systems and report what they find. Cathedral, a cybersecurity startup founded months ago by four former staffers of the Department of Government Efficiency, carried out the exercise and disclosed it on October 26, 2026.

The way in: ordinary business software

The entry point was not Anthropic's model. Cathedral says the agents exploited unglamorous third-party software every company runs in the background: tools for administration, sales, HR and IT. Once inside, the agents reached employee compensation details, internal communications, meeting transcripts, salary and job-offer information, and Slack messages discussing Anthropic's advanced models, according to Cathedral's draft report.

That detail is the point for enterprise security teams: the agents did not beat the AI, they beat the procurement stack around it. Every company that buys software from vendors now has the same exposed surface, whether or not it builds models. Anthropic itself has reported similar offensive capability before, saying in November 2025 that Chinese state-sponsored hackers used its Claude Code coding tool to attack roughly 30 organizations, and disclosing in July that Claude models hacked real organizations in three incidents during its own tests.

Anthropic disputes the severity

Anthropic confirmed that a limited set of sales, HR and IT systems was accessed, but says customer conversations, Claude responses, source code and model weights were never reached. Model weights, the trained parameters of an AI model, are considered a company's most sensitive AI asset.

Our monitoring flagged some of Cathedral's activity before they reported it to us, we shut off their access the same day," an Anthropic representative said.

The two sides disagree on more than the timeline. Anthropic says Cathedral overstated the national security implications of the findings and did not disclose all of its testing. Cathedral says data access was kept to the absolute minimum in line with security research best practices. Neither account is independently settled, and Cathedral's report is still a draft.

Which vendors supplied the affected software, and which specific vulnerabilities the agents exploited, have not been named.

The post AI Agents Reached Anthropic's Business Systems in Under 12 Hours appeared first on Techreport.