AI Vulnerability to 'Many-Shot Jailbreaking' Exposed by Anthropic