The most valuable lesson from this hack isn't what Cathedral found at Anthropic—it's what they ignored before looking there.
The Summary
- Cathedral, a military-focused AI security startup run by former DOGE staffers, claims it breached Anthropic's internal systems in under 12 hours, accessing meeting transcripts, salary data, and Slack messages about AI models
- The breach happened through third-party business software vulnerabilities, not Anthropic's core AI systems—which Cathedral failed to penetrate for "multiple days"
- Cathedral just raised $160 million at a $1.4 billion valuation from A16z and Sequoia, positioning itself as the defense contractor for the agent economy
The Signal
Cathedral spent days trying to break into Anthropic's AI infrastructure and failed. Then they pivoted to the boring stuff—the Slack integrations, the HR software, the meeting schedulers—and got in before lunch. That failure-then-pivot tells you more about AI security than the breach itself.
The companies building frontier AI models have fortified their core systems. They've hired the best security talent, built custom infrastructure, and assumed nation-state adversaries from day one. Anthropic's moat held. Cathedral couldn't touch it.
"The fact that weaknesses in ordinary enterprise software can compromise the infrastructure of even the most advanced organizations should be a wake-up call."
But every AI lab still runs on the same third-party plumbing as everyone else. The same calendar apps. The same project management tools. The same business software that gets installed because it integrates with everything and makes life easier. That's the gap. Not model weights or training data, but the 47 SaaS tools in the stack that someone in ops approved without a security review.
Here's what makes this timing interesting:
- Cathedral closed a $160M Series A in July at $1.4B valuation
- Two months later, they're publicizing a hack of one of the three companies that matter most in AI
- They're positioning as military-grade security for AI infrastructure
- Anthropic confirmed this was a sanctioned penetration test through their bug bounty program
This is marketing. Cathedral isn't warning the industry out of civic duty. They're demonstrating capability to prospective customers—AI labs, defense contractors, anyone building agents that will touch sensitive systems. The message: "We got into Anthropic. We can protect you."
The substantive point stands, though. As AI agents proliferate, the attack surface isn't the models themselves. It's the connective tissue. Every calendar invite an agent reads. Every Slack channel it monitors. Every API it calls to book travel or file expenses. These aren't theoretical vectors. Cathedral just proved you can map an organization's AI development roadmap by reading meeting transcripts in compromised business tools.
The Implication
If you're building with AI agents, audit your stack. Not your model security—your Zapier connections. The tools your agents use to access email, calendars, project management, and internal comms are now infrastructure. Treat them like it.
For Cathedral, this is the pitch deck brought to life. Former government efficiency crusaders turned cybersecurity founders, funded by the biggest names in tech, hacking the AI company everyone's watching. Whether they found anything valuable matters less than proving they could get in. In the agent economy, that proof is the product.