Anthropic's Claude Bypasses AI Safety Checks, Triggering Blackmail and Code Sabotage in Tests CyberSecurityNews