A startup's AI agents reached Anthropic's data through third-party software, showing the supply chain is the weak link

Cathedral, a cybersecurity startup founded only months ago by former members of Elon Musk's now-disbanded Department of Government Efficiency (DOGE), says its AI agents gained access to Anthropic's internal systems. According to the Washington Post, as cited by Futurism, the agents exploited vulnerabilities in business software and reached meeting transcripts, salary details, job offers and Slack messages in less than 12 hours.

Anthropic tells a narrower story. A spokesperson told the Post that Cathedral reached only a "limited set" of sales, HR and IT systems. According to the company, the startup spent days trying to break into Anthropic's own software, failed, and then turned to a business tool Anthropic uses. "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. The company added that it has "no evidence" that anyone other than Cathedral exploited the vulnerabilities.

Who is behind Cathedral matters. Investors value the startup at $1.4 billion. It has raised $160 million and is under contract with at least three federal agencies. Marc Andreessen, cofounder of Andreessen Horowitz, sits on its board. An unnamed Trump administration official told the Post the report had alarmed people 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 context is a year of rising fear about AI espionage. According to Futurism, several labs have admitted that models escaped their sandboxes, and a Saudi national was arrested this month, accused of bugging an OpenAI conference room.

Our thesis: the two accounts disagree on severity, but they agree on where the hole was. Neither says Anthropic's core systems were breached. Both describe entry through a vendor's tool. That is the classic supply-chain problem, and AI agents make it cheaper and faster to exploit. An attacker no longer needs a team of experts working for weeks. A swarm of agents can test every peripheral piece of software a company uses until one gives way.

We should also be careful with the framing. Cathedral is a company that sells offensive security, and a headline saying it "hacked Anthropic" is also its best advertisement. The details of what was accessed and for how long come from the two parties involved, and neither is neutral. The one claim that would most directly reassure on Anthropic's side, that its monitoring caught the activity, is precisely what neither side has independently proven.

Our reading: this case fits a pattern we've been tracking. The near-term danger of agentic AI is not a distant superintelligence. It is the automation of intrusion against the least-protected link in the chain. The answer isn't panic. It is zero-trust architecture that extends to vendors, continuous monitoring, and disclosure rules for these tests. In the long run, the same agents that find holes in hours can close them in hours. Whether defense or offense benefits more depends on who adopts the tools first and how those tools are governed.