Cathedral, a cybersecurity start-up only months old, has told Anthropic that its researchers penetrated sensitive internal data and messaging systems at the company. The Washington Post reports this and cites a draft of Cathedral’s report that the newsroom was shown. By the paper’s account, the access happened under a program in which Anthropic invites outside experts to test its own systems.
The report rests on a single investigation by The Washington Post, bylined by Elizabeth Dwoskin and Gerrit De Vynck and marked as an exclusive. No second, independent version exists so far. The article sits behind a paywall; the headline, the deck and the first paragraph are readable, and that is the only material behind what you read here.
Cathedral is not an unknown name
The start-up was founded by former DOGE staffers, the paper says. That a firm of that kind can open up the systems of one of the leading AI labs is something the Post places in the context of mounting concern about security in the global AI race.
The genre is not new. In September, three researchers used Claude Opus 5 to find their way into OpenAI’s monorepo, starting from a doctored image in the OpenAI forum, and it took them under 72 hours. The labs are now both things at once: tool supplier for attackers and target.
What matters here is the timing
Anthropic filed its S-1 with the SEC on Monday, and even the draft prospectus named serious security incidents as a risk whose damages could exceed the company’s insurance coverage. On the same day, a report appears saying somebody got into internal messaging systems. That it happened during an invited test is the decisive difference from a real attack — and also the reason any of it became public at all. What exactly Cathedral achieved sits in a report I have not read. Once more of it is public, or a second newsroom follows up, I will come back to it.