By Hamza Faisal

A months-old cybersecurity startup founded by former members of the US Department of Government Efficiency says its AI agents found their way into Anthropic's internal systems in under 12 hours, according to a draft report reviewed by The Washington Post and reported on Monday.

Cathedral, the startup behind the test, said its agents exploited weaknesses in third-party administrative and sales software used by Anthropic, rather than breaking into the AI lab's own code. From there, the researchers said they reached meeting transcripts, employee salary and job-offer details, and internal Slack discussions about Anthropic's most advanced AI models.

The test ran under Anthropic's bug bounty programme, which invites outside researchers to probe its systems and report flaws. Cathedral said it disclosed its findings to Anthropic earlier this month and told the affected software vendors, whose names have not been made public.

A 12-hour path through third-party software

According to Cathedral's draft report, the agents identified routes into Anthropic's corporate systems in less than 12 hours by exploiting flaws in third-party business applications. The company said its researchers kept data access to a minimum.

Cathedral was founded this year by former DOGE staffers including Gavin Kliger, who served as the Pentagon's chief data officer until April. The startup raised $160 million at a $1.4 billion valuation in July, with backing from Andreessen Horowitz and Sequoia Capital, and contracts with three federal agencies, according to reports.

The background matters: Kliger was involved in the Pentagon's high-profile falling-out with Anthropic earlier this year, after the department restricted the company's models in federal systems following Anthropic's refusal to allow their use for autonomous weapons or mass domestic surveillance.

Anthropic disputes the scale of the breach

Anthropic confirmed that Cathedral reached a limited set of its sales, human resources and IT systems, including business contact information belonging to some customers and partners and basic user account records. But it drew a firm line around its core assets: customer conversations with Claude, model responses, proprietary source code and model weights were not accessed, the company said.

Anthropic also said its own monitoring flagged some of Cathedral's activity before the startup reported it, and that it shut off the researchers' access the same day. The affected software was disabled until vendors shipped fixes, and federal officials were notified. The company said it found no evidence that anyone outside Cathedral had exploited the same weaknesses.

The two sides disagree on the framing. Anthropic accused Cathedral of overstating the national security implications and of failing to disclose the full extent of its testing under the bug bounty programme's rules. It also noted that the researchers had spent several days unsuccessfully trying to break into its own software directly before turning to the third-party applications.

Why Washington is paying attention

The disclosure lands amid a wider debate about how well leading AI labs can control increasingly autonomous systems, just days before US midterm elections on November 3. A senior Trump administration official told the Post that if a newly founded startup could get in, then foreign adversaries may not be far behind.

For the AI industry, the incident also underlines an uncomfortable point: the softest target may not be the frontier model itself, but the ordinary business software wrapped around it. An AI agent that can chain their weaknesses together does not need to defeat the model at all.

Anthropic says mitigations are now in place. Whether other labs will face the same kind of test, and publish the results, is now the open question.