TechBroz · techbroz.com44 incidents on record · updated as they happen
← Hall of Shame

Rank #16 of 44·Security·Oct 2026·Cathedral's AI hacking agents (security startup, testing via Anthropic's bug bounty programme)

A startup's AI agents got into Anthropic's sales, HR and IT systems through third-party software, and the two sides dispute how far

A server room, for illustration; not Anthropic's systems (edited).
A server room, for illustration; not Anthropic's systems (edited). Photo: Carl Lender, CC BY 2.0, via Wikimedia Commons; cropped, text overlaid.

Disputed Cathedral says it took under 12 hours and reached salaries, meeting transcripts and Slack messages. Anthropic confirms a limited set of sales, HR and IT systems, says attempts ran over several days, and says Cathedral reported only one of two flaws.

“It didn't pick the front-door lock. It used the side door the HR software left open.”

Incident reportSecurityDisputedThird-party admin and sales software connected to Anthropic
Agent / toolCathedral's AI hacking agents (security startup, testing via Anthropic's bug bounty programme)
CompanyAnthropic (target); Cathedral (tester)
DamageRoutes to salary data, meeting transcripts and Slack messages, per Cathedral; Anthropic says a limited set of systems
Time to detectAnthropic says its monitoring flagged some activity before Cathedral reported it, and it cut access the same day
StatusAccess cut and affected software disabled until vendors fixed it; timing and scope disputed
Chaos8.0 / 10
  • The way in was third-party admin and sales software, not Claude
  • Anthropic says Claude chats, source code and model weights were not reached
  • Anthropic says the attempts ran for days and one of two flaws went unreported; Cathedral says it followed best practice

Anthropic pays outsiders to find holes in its systems. A months-old startup pointed AI agents at it and came back with routes to salaries and Slack messages.

What happened

Cathedral, founded this year by four former DOGE staffers, tested Anthropic through its bug bounty programme. Its draft report says its AI agents, which carry out multi-step tasks without a person approving each move, got in not through Claude but through weaknesses in third-party admin and sales software that many companies use. They found routes to meeting transcripts, salary and job-offer details, and Slack messages about Anthropic's advanced models.

The damage

Anthropic confirms the agents reached a limited set of sales, HR and IT systems holding business contact details and basic account records. It says customer chats with Claude, source code and model weights were never reached, and it has no sign anyone else used the flaws. Its monitoring flagged some of the activity before Cathedral reported it, it cut access the same day and it disabled the affected software until the unnamed vendors fixed it.

How it was caught, and what happened next

The two sides don't agree. Cathedral says it took under 12 hours. Anthropic red-team lead Logan Graham says there were several days of attempts and that Cathedral found two flaws but reported only one. Anthropic says the startup broke the programme's rules; Cathedral says it kept data access to the minimum.

Lesson for builders

Your weakest link is often the business software you plugged in. Know every app that can reach HR and chat data.

Fail card: A startup's AI agents got into Anthropic's sales, HR and IT systems through third-party software, and the two sides dispute how far
The fail card

This story in one shareable card. Tap to open it full size.

Share on X All fail cards →

More fails

Get TechBroz Weekly free → All 44 fails