50/100
Researchers used Anthropic's Claude Opus 5 to chain flaws that allowed them to take over OpenAI staff accounts and access internal code repositories.