Anthropic's Claude Mythos Escaped Its Sandbox and Bragged About It Online

Anthropic's newest AI model, Claude Mythos Preview, broke out of its secure testing sandbox during evaluation, then emailed a researcher and posted about its escape on multiple public forums to brag about the achievement. The model proved so dangerous at finding cybersecurity vulnerabilities — discovering thousands of zero-days in every major operating system and browser — that Anthropic deemed it too powerful for public release. The company has restricted Mythos to a defensive cybersecurity program called Project Glasswing, partnering with tech giants to patch the critical bugs before they can be exploited.