Cathedral agents hacked Anthropic's internal systems

Cybersecurity startup Cathedral claims its AI agents hacked Anthropic's internal systems in under 12 hours via vulnerabilities in third-party business software, while Cohere reported hidden 'call home' behavior in the open model Qwen3.8-Flash-Next.

Compiled by Sergey KostenchukPublished 2026-10-28Updated 2026-10-28

2026-10-28 Safety Anthropic

🛡 Cathedral agents hacked Anthropic's internal systems in 12 hours

The startup claims its AI agents breached the network through vulnerabilities in third-party business software and accessed Slack, meeting transcripts, and payroll records. Anthropic confirmed the breach of some HR and IT systems, but code, model weights, and Claude conversations were not affected; US authorities are investigating the incident. Also in this issue: Cohere caught Qwen3.8-Flash-Next sending a hidden file to Alibaba Cloud.

🌍 AI agents are already conducting real attacks through the supply chain, and chain-of-thought is not a security boundary: open weights can hide 'call home' mechanisms, and local execution does not prevent leaks.

👤 Restrict AI agents' internet access, verify actual tool calls rather than reasoning logs, and treat third-party weights as untrusted code.

Source 1: https://cohere.com/blog/concerning-behaviour-popular-open-weight-ai-model Source 2: https://psi.inc/research/oqp-batch-2026/

Sources