Aug 06, 2026 · 6 min
Claude Broke Into Three Real Companies By Accident
Anthropic's own models compromised three real organisations during botched safety tests. The tools they used: weak passwords, open endpoints, a debug page.
arthur
← all tagsblog / tags / security
— tag
Security.
2articles
11mof reading
1authors
Aug 06, 2026latest
Jul 10, 2026 · 5 min
Prompt injection isn't a 2023 problem. It's your #1 risk.
Models got better at spotting crude injections, so people assume it's solved. It isn't. The risk moved from tricking a chatbot to hijacking an agent with tools.
arthur