Cathedral's breach of Anthropic Slack

Cathedral, a security startup that had been operating roughly three months before the incident, accessed Anthropic's Slack workspace by exploiting vulnerabilities in third-party business tools rather than attacking Anthropic's core systems directly. Anthropic confirmed Cathedral reached a limited set of sales, HR, and IT systems, and said its own monitoring flagged Cathedral's activity before they self-reported, with access shut off the same day. Federal officials were briefed. Logan Graham of Anthropic corrected press coverage that described a 12-hour attack, stating the intrusion spanned several days and exploited two third-party vulnerabilities. Graham said nothing reached Anthropic's most sensitive systems, including model weights or Claude conversations, and characterized the incident as outside Anthropic's safe harbor policy because it involved exploiting third parties and accessing data rather than a standard bug bounty finding. A separate account alleged Cathedral had achieved full administrator access to the Slack workspace, enabling them to see every message and log in as arbitrary users, and described the workspace as containing "literally every single research secret," characterizing Anthropic as massively underplaying the breach. Critics, as characterized by one source, argue both Anthropic and OpenAI keep their most sensitive research secrets on Slack, making them vulnerable to foreign intelligence services, and that both companies build primarily on Slack because, as one source put it, no one has the time to figure out something better. Graham described the incident as an example of what he called the future of cybersecurity: adversaries deploying AI agents continuously to probe dependencies, with models roughly half to one generation behind the frontier capable of conducting such attacks, and called for continuous AI-assisted red teaming and redesigning software to reduce attack surfaces. Brundage argued that no legislator has made AI security, distinct from AI safety, a top priority.