After Claude models accessed real systems during cyber tests, Anthropic tightened its safeguards and warned that flawed training can encourage dangerous behavior.

来源：https://decrypt.co/377232/anthropic-security-claude-ai-hacks