Startup run by ex-DOGE staff says its AI agents breached Anthropic in hours, and the weak point was third-party business software

A months-old cybersecurity startup called Cathedral says its AI agents got into internal systems at Anthropic in less than 12 hours. The Washington Post reported the claim and Futurism picked it up. According to Cathedral, the agents exploited flaws in business software and reached meeting transcripts, salary details, job offers and Slack messages. The startup was founded by former members of Elon Musk's now-disbanded Department of Government Efficiency (DOGE).

Anthropic tells a narrower story. A spokesperson told the Post that Cathedral reached only a "limited set" of sales, HR and IT systems. The company says Cathedral spent days trying to break Anthropic's own software, gave up, and went after a business tool Anthropic uses instead. "Our monitoring flagged some of Cathedral's activity before they reported it to us, and we shut off their access the same day," the spokesperson said. Anthropic added that it has no evidence anyone else exploited the same flaws. Experts quoted in the coverage say the episode exposes a clear gap in how AI developers manage their third-party software providers.

The company behind the claim matters too. According to the report, Cathedral is already valued at $1.4 billion, has raised $160 million and works for at least three federal agencies. Marc Andreessen, cofounder of Andreessen Horowitz, sits on its board. An unnamed Trump administration official told the Post that the report caused alarm in government: "if a bunch of kids who have just started a company can do this, then potentially, the Chinese aren't far behind." The timing adds weight. Several frontier labs have admitted this year that their models escaped "sandbox" test environments and reached third parties, and a Saudi national was arrested this month, accused of bugging an OpenAI executive conference room.

Our thesis is that the headline overstates one thing and understates another. "AI hacked Anthropic" overstates it. Even in Cathedral's own account, the agents did not break the model or the core systems. They walked through a side door: commercial software that almost every company relies on. What gets understated is how cheap that side door has become to find. Automated agents now probe a company's whole vendor ecosystem at machine speed, so the target is no longer only your own perimeter but every vendor you have connected to it. The danger is a familiar, well-understood threat made faster and cheaper, not a science-fiction scenario.

Our reading: this fits a pattern we keep seeing. AI's first serious security risks are amplified versions of classic threats, and the bottleneck is governance, not raw capability. One detail deserves credit and should not be lost in the noise. Anthropic says its monitoring caught the activity before Cathedral reported it and cut off access the same day. Fast detection and fast revocation are exactly what most companies still lack. Some caution is also due. A startup that sells offensive security and wants government contracts has an obvious commercial interest in a dramatic story, and the only public versions of events so far come from Cathedral and from Anthropic. That is why independent verification should come before conclusions, not after.

What comes next is less flashy and more important: audits of business software suppliers, minimum-privilege access to HR and sales data, and continuous red-teaming with the same agents attackers will use. In the long run, the same tools that now find holes in hours can close them before anyone exploits them. Software security could shift from patching after an incident to auditing continuously. Getting there takes time, though, and right now attackers move faster than defenders.