The Deep Dive reports access to Anthropic business systems during authorized testing

The Washington Post reported that researchers had penetrated sensitive internal data and messaging systems at an artificial intelligence company, and said its account was based on a draft Cathedral report it reviewed.[1] The Post also reported that Cathedral recently told Anthropic its researchers had penetrated systems.[1] Techmeme’s headline characterized the incident as Cathedral claiming a breach of Anthropic’s internal systems in under 12 hours.[2] Its excerpt said the breach involved AI agents.[2]

The Deep Dive reported that Cathedral’s access to Anthropic’s business systems happened during authorized security testing.[3] That qualification matters when interpreting the word “breach”: the supplied reporting describes the access as occurring in a testing context, while the Washington Post account describes penetration of sensitive internal data and messaging systems.[1][3] The evidence does not give a detailed account of the authorization, the testing boundaries, or whether the descriptions cover identical systems. For security and risk teams, those distinctions are important questions to resolve before drawing conclusions about exposure or control effectiveness.

The Washington Post described Cathedral as a months-old cybersecurity startup founded by former members of the U.S. DOGE Service.[1] The Deep Dive reported that a co-founder was a former Pentagon official involved in the U.S. government’s legal dispute with Anthropic.[3] It also reported that Cathedral raised $160 million earlier this year at a $1.4 billion valuation.[3] Reuters reported that Andreessen Horowitz and Sequoia Capital led the funding round and took board seats.[4] Reuters said Cathedral aims to use AI to expand U.S. military cyber capabilities; the Washington Post reported that its AI-powered cybersecurity tools serve government and corporate clients.[4][1]

The Washington Post situated the reported penetration amid mounting concerns about security in the global AI race.[1] It reported that Cathedral offers AI-powered cybersecurity tools to U.S. government and corporate clients.[1] That combination makes the account relevant to organizations weighing AI-enabled security testing: the reporting presents both a penetration claim and an authorized-testing qualification.[1][3] The Washington Post published its report on October 26, 2026.[1]