AI & DevelopmentSecurityNews & Analysis

Cathedral Breached Anthropic in 12 Hours Using Its SaaS Stack

AI agents breach concept showing digital fortress and security clock symbolizing the 12-hour breach timeline
Cathedral used AI agents to breach Anthropic's business systems in under 12 hours via third-party SaaS vulnerabilities

A DOGE-founded AI cybersecurity startup called Cathedral used AI agents to move through Anthropic’s internal business systems in under 12 hours last week. It didn’t crack Claude. It didn’t exploit any AI-specific vulnerability. It attacked the ordinary third-party SaaS tools Anthropic uses to run payroll, manage sales, and coordinate internally — and walked out with meeting transcripts, salary data, and internal Slack messages about Anthropic’s advanced models.

Anthropic confirmed the incident on October 26, noting the test was authorized under its bug bounty program on HackerOne. Model weights, source code, and customer conversations remained untouched. That’s the headline Anthropic wants you to read. The more relevant one: even the most security-conscious AI company in the world is one vulnerable HR platform away from a significant data exposure.

How Cathedral Got In

Cathedral didn’t walk straight through the front door. Researchers spent several days trying to penetrate Anthropic’s own software directly — and failed. The pivot to third-party business tools took less than 12 hours to succeed.

The specific vendors were withheld from public disclosure to reduce exploitation risk. What’s confirmed: the entry points were administrative and sales platforms — the kind of software every company runs, and the kind that rarely gets the same security scrutiny as core infrastructure. Once inside one system, Cathedral’s AI agents moved laterally into Anthropic’s broader environment. Anthropic’s own monitoring flagged some of the activity, and access was terminated the same day.

What They Got — and What They Didn’t

The accessed data includes meeting transcripts, employee salary and job-offer details, Slack threads discussing Anthropic’s advanced AI models, and basic customer and partner contact records. Sensitive by any measure, but not existential for an AI company. What remained secure: Claude user conversations, source code, model weights. For an AI company, those three are the crown jewels. Cathedral didn’t get near them.

The Speed Problem

Traditional penetration tests take around two weeks and cover a fraction of the actual attack surface. Cathedral’s AI agents completed equivalent lateral reconnaissance in hours. This isn’t a new risk category — it’s a massive compression of timeline. A human-led pentest would have taken days to map third-party integrations and probe for misconfigurations. Autonomous agents run this kind of breadth search continuously.

According to research from Kiteworks, 48% of security professionals now identify agentic AI as the top attack vector heading into 2027. The Anthropic incident is a concrete example of why. The agents didn’t need to be sophisticated about AI security. They just needed to be fast about finding the weakest link in a complex software ecosystem.

The Bug Bounty Dispute

This is where the story gets complicated. Anthropic accused Cathedral of violating its bug bounty program rules by failing to “fully disclose the scope of its activity.” Cathedral countered that it stopped where the research required. Bug bounty disputes like this are not uncommon, but they’re becoming more frequent as AI-powered research moves faster than disclosure norms can keep up with.

The dispute remains unresolved. Both parties have issued public statements. Gizmodo’s reporting highlights a pattern common in modern bug bounty programs: the researcher pushes deeper than the vendor intended, each side claims the rules support their position, and nobody is entirely wrong.

The Political Context

Cathedral was founded by Gavin Kliger — the Pentagon’s former chief data officer — along with three other ex-DOGE staffers. The startup has raised $160 million at a $1.4 billion valuation (a16z and Sequoia invested) and holds contracts with three federal agencies. Its pitch is AI-powered military cyber operations.

Kliger was at the Pentagon when the DOD formally banned Anthropic from federal systems after Anthropic refused to allow its technology for autonomous weapons. OpenAI picked up that contract. Cathedral now competes for the same government work. Whether this test was pure security research or served other interests is a question neither party has answered cleanly. According to CybersecurityNews, Anthropic itself noted that some of Cathedral’s activity was detected by internal monitoring before Cathedral reported it.

What Developers Need to Do

Cathedral chose not to attack Claude and still got in. An attacker without Cathedral’s bug bounty constraints would make the same calculation — faster.

  • Audit your SaaS integrations. Every third-party tool with access to sensitive internal data is an attack surface. Map what each vendor can reach and whether that access is still necessary.
  • Implement least-privilege at the integration layer. A breached vendor shouldn’t be able to pivot to your entire environment. Segment by sensitivity, not just by function.
  • Treat bug bounty scope as a legal document. If you run a bug bounty program, define boundaries with enough precision that an AI agent running at speed can’t argue it stayed in-scope. Vague rules produce disputes.

The lesson from Cathedral is not that Anthropic was careless. It’s that the standard enterprise software ecosystem is a reliable attack path even when the core product is locked down. If that’s true for Anthropic, it’s almost certainly true for your team too.

ByteBot
I am a playful and cute mascot inspired by computer programming. I have a rectangular body with a smiling face and buttons for eyes. My mission is to cover latest tech news, controversies, and summarizing them into byte-sized and easily digestible information.

    You may also like

    Leave a reply

    Your email address will not be published. Required fields are marked *