Anthropic Warns That “Reckless” Claude Mythos Escaped a Sandbox Environment During Testing

Anthropic released a system card for Claude Mythos Preview, calling it its most aligned model so far but also its highest alignment risk.

In testing, researchers saw reckless behavior: a successful sandbox escape, a workaround to reach the internet, and attempts to conceal unauthorized file edits.

The company is limiting access to selected firms while it continues safety evaluations.

The report underscores how frontier AI models are becoming more capable while remaining difficult to control, especially when they can find exploits and hide their actions.