The AI offence-defence debate is asking the wrong question On 7 April 2026, Anthropic announced that an unreleased frontier model, Claude Mythos Preview, had found thousands of zero-day vulnerabilities – on its own, across every major Gil Baram July 7, 2026