Startup says its AI agents reached Anthropic's internal systems through third-party software

Cathedral says its agents found paths into Anthropic's business systems in under 12 hours. Anthropic confirms limited access but disputes the account.

27 October 2026

What happened

Cathedral is a cybersecurity startup founded by former DOGE staffers. It says AI agents it deployed found routes into Anthropic's internal business systems in less than 12 hours. The account comes from a draft Cathedral report reviewed by The Washington Post, which published on 26 October. Gizmodo and Fox News relayed the report.

According to both outlets, the agents did not attack Claude. They exploited weaknesses in third-party business software that Anthropic uses. The work took place within Anthropic's bug bounty program, which pays outside researchers who find vulnerabilities.

What Cathedral says it reached

Gizmodo reports that, according to Cathedral's report, the agents found paths to meeting transcripts, salary and job-offer information, and Slack messages. Fox News describes the Slack material as messages about advanced AI models. Cathedral's report argues that weaknesses in ordinary enterprise software can compromise even the most advanced organisations.

The date of the intrusion and the date Cathedral reported it to Anthropic have not been published.

Anthropic's response

According to Gizmodo, Anthropic confirmed that the researchers reached a "limited set" of sales, HR and IT systems. These included business contact details for customers and partners and some basic user account records. Anthropic said the researchers did not access Claude users' conversations, source code or model weights. Fox News adds model responses to that list. Gizmodo also reports Anthropic said its own researchers had spent several days trying, without success, to breach its software directly.

According to Fox News, Anthropic also said:

it detected some of Cathedral's activity before the startup disclosed its findings
it shut down the affected access points
it notified federal officials
it has no evidence anyone else exploited the flaws

Gizmodo's article does not mention a notification to federal officials.

Points in dispute

The two sides describe the scope differently. Cathedral emphasises transcripts, pay data and internal messages. Anthropic describes contact details and basic account records in a limited set of systems.

Anthropic also disputes how Cathedral went about the testing. Fox News reports Anthropic said the startup did not fully comply with its security-testing rules. Gizmodo says Anthropic accused Cathedral of breaking the program's rules by not disclosing the full extent of its activity.

There is also disagreement about how serious the findings are. An unnamed senior Trump administration official told The Washington Post that the episode raised national security concerns, including that Chinese hackers could exploit similar weaknesses. Fox News reports that Anthropic disputed Cathedral's account of the national security risks.

About Cathedral

Gizmodo reports the company was founded by Gavin Kliger, Luke Farritor, Marko Elez and Jack Stein. Kliger previously served as the Pentagon's chief data officer and was reportedly involved in the Pentagon's dispute with Anthropic. Gizmodo says Cathedral is valued at US$1.4 billion and has raised US$160 million from investors including Andreessen Horowitz and Sequoia Capital.

Context

The report came out on the same day Anthropic made its IPO prospectus public, according to Fox Business.

This case is different from most recent agent incidents. In those cases, models under test acted beyond what their operators intended. Here, outside researchers pointed agents at a target under a bug bounty program. The questions are about the agents' speed and reach, and whether the researchers stayed within the program's rules.

The Washington Post article was not accessible for this report, so details attributed to it come from Gizmodo and Fox News.