● Who: Cathedral, a cybersecurity startup
● Time to access: under 12 hours
● Attack path: third-party business software, not Anthropic products
● Anthropic says reached: a limited set of sales, HR and IT systems
● Anthropic says NOT reached: customer conversations, model weights, source code
Cathedral, a cybersecurity startup, says its researchers used AI agents to reach sensitive internal Anthropic business systems in less than 12 hours, exploiting weaknesses in third-party business software rather than Anthropic's own products. The Washington Post published the investigation on Monday 26 October.
The systems named include internal meeting information, Slack discussions, HR records and other administrative data. Anthropic's account is narrower: it says the access covered a limited set of sales, HR and IT systems, and specifically did not include customer conversations, model weights or source code. It disabled the affected connections and later restored them.
Anthropic also disputes the broader national security framing and says the researchers did not fully follow its vulnerability disclosure rules. The detail worth noting for anyone running agents is the attack path: not the model, but the third-party SaaS connected to it.
Sources
Tech Startups daily brief.