AI startup Cathedral breaches Anthropic network in under 12 hours

Artificial intelligence startup Cathedral has breached the internal infrastructure of Anthropic in less than 12 hours using autonomous AI agents. The incident occurred as part of a Bug Bounty programme hosted on the HackerOne platform, where security researchers are tasked with identifying system flaws. However, the case has highlighted once again how vulnerable even high-security tech giants can be when relying on third-party software.

How the attack was carried out

Cathedral researchers initially spent several days trying to penetrate Anthropic's systems using traditional human methods, without success. They then introduced custom autonomous AI agents into the process. The AI quickly identified weaknesses in third-party business software and secured access to the company's internal network in under 12 hours.

Once inside the network, the agents gained access to internal meeting transcripts, employee salary details, job offers, and Slack communications. In an official statement, Anthropic emphasised that the attackers did not access Claude's AI model weights, source code, or private user chats. However, partial data from the company's sales, human resources, and IT systems was exposed, including contact information for certain partners and clients.

Rule violations and pre-IPO challenges

Anthropic manages a two-part Bug Bounty programme through HackerOne. One track focuses on identifying technical vulnerabilities, while the second—an invite-only scheme—rewards researchers for bypassing Claude's safety guardrails (jailbreaking). Cathedral submitted its findings under this programme.

Nevertheless, Anthropic stated that its internal monitoring systems flagged Cathedral's activity before the official report was submitted, and access was revoked on the same day. The company also accused Cathedral of breaching programme rules, claiming the startup failed to disclose the full scope of its actions in a timely manner. It remains unclear whether Anthropic will pay the researchers the bounty.

The timing of the incident is particularly sensitive for Anthropic. Valued at $1.4 billion, the company is preparing for an initial public offering (IPO). Such data exposures, even when triggered by white-hat security testing, present an additional risk factor for investors. Cathedral's report noted that flaws in standard corporate software should serve as a wake-up call for every organisation.

Who is behind Cathedral and what experts say

Cathedral is no ordinary startup. Founded by former members of the US Department of Government Efficiency (DOGE), the firm focuses on developing offensive and defensive cyber operations for US military and government entities. Its founders include former Pentagon Chief Data Officer Gavin Kliger, machine learning researcher Luke Farritor, former SpaceX employee Marko Elez, and former Salem Partners investor Jack Stein. A senior US administration official noted that if a newly formed company can achieve this so quickly, potential adversaries will not be far behind.

The incident triggered widespread discussion across the cybersecurity community. Jacob Krell, a representative at Suzu Labs, noted that third-party vendors always carry risk: "A company's security level is limited by the weakest partner you integrate, and internal engineering cannot change that."

Darin Fredde, an expert at Ridge Security, acknowledged Anthropic's rapid response but added: "Real security is independent validation of attack paths. We shouldn't confuse a successful patch with proven security, or we're just locking the front door hoping no one finds the spare key."

What this means for the industry

The advancement of autonomous AI agents elevates the speed and efficiency of cyberattacks to a new level. Where uncovering flaws in complex corporate infrastructure previously required weeks or months, AI agents can now complete the task in hours.

For organisations, including technology firms and financial institutions in Georgia and beyond, this case provides a clear lesson that safeguarding one's own core systems is no longer enough. Continuous monitoring of third-party software and partner services is essential, as modern AI tools will locate even the smallest oversight in minimal time.