Anthropic's latest frontier AI model, Claude Mythos Preview, is so adept at finding software vulnerabilities that the lab is holding it back to allow companies and institutions to proactively patch their products against the 'thousands' of bugs it has already uncovered.